The Technical Cooperation Program Australia - Canada - New Zealand - United Kingdom - United States of America TTCP TECHNICAL REPORT JSA Action Group 14 Complex Adaptive Systems for Defence Progress Report Causal & Influence Networks in Complex Systems May 2010 DOC – AG14 - C&IN #4-2010 Correspondence: Anne-Marie.Grisogono@defence.gov.au This report contains Information which is provided in confidence to the Governments of Australia, Canada, New Zealand, the United Kingdom and the United States of America under the auspices of The Technical Cooperation Program (TTCP). The Information contained herein may be used and disseminated for national Defense Purposes within the recipient Governments and their national defense Contractors. The recipient Governments will ensure that any other use or disclosure of the Information is made only with the prior written consent of each of the above Governments. CONTENTS Introduction ................................................................................................................. 3 Causation in Complex Systems ................................................................................ 6 Causal and Influence Networks .............................................................................. 6 Challenges to Understanding C&INs ..................................................................... 8 Decision Support in Complex Situations ............................................................. 11 Existing C&IN Approaches – capabilities and limitations ................................ 15 Bayesian techniques ................................................................................... 15 System Dynamics ....................................................................................... 15 Network theory .......................................................................................... 16 Simulation ................................................................................................... 17 Non-Linear Dynamical Systems .............................................................. 17 Strategy ....................................................................................................................... 18 State-of-the-Art review ........................................................................................... 18 Novel CAS Approaches .......................................................................................... 20 Promising CAS Concepts ....................................................................................... 20 Multi-scale Intent ....................................................................................... 21 Attractors ..................................................................................................... 22 What is known about interacting adaptive processes? ........................ 24 Implementation: The C&IN Program ................................................................... 25 Scoping Paper and Open Questions ..................................................................... 25 Workshop Series ...................................................................................................... 26 Deliverables and Forward Workplan ................................................................... 28 References................................................................................................................... 31 Annex A: Essential and Desirable Requirements for a C&IN Support Tool for Complex Operations .......................................................................................... 34 Annex B: International Collaborative Research Program ................................. 38 PAGE 2 OF 54 Introduction Context - and why this is a progress report rather than a final report. From its inception, AG14 recognised two fundamental research challenges at the root of most complexity problems: how structure arises and develops, and how causation works in complex systems. Research thrusts were established in both, as well as in several application areas, as described in the Final Report [1] which this document accompanies. The outcomes of the first thrust, a Conceptual Framework for Adaptation, presented in an accompanying Synthesis Report [2], have been applied extensively in much of AG14’s work. The second thrust, addressed here, turned out to be much more challenging, not surprisingly, since it is in a sense the problem of complexity, and subsumes the issues surrounding how structure arises and develops. What we have learned within the first thrust is therefore relevant to and helpful for addressing the second, as will be discussed below. In initial discussions with complex systems theorists and practitioners (eg at the Santa Fe Institute, within CSIRO, and at various universities) about causation in complex systems, a frequently encountered response was that the scope of what we were seeking to address was far too sweeping and grandiose, and we were encouraged to focus our attention on a well-defined problem that could be worked on in more detail. But AG14 was already doing this in its application thrusts, as well as individual members doing so in the normal course of their defence work, and the challenges of understanding complex causation were recurring themes which limited the extent to which each selected problem could be satisfactorily addressed. So we persisted, against well-intentioned and well-informed advice, but recognising the magnitude of the task, and that AG14 would therefore need to motivate support and cooperation from many quarters in order to make any headway. As a logical step to doing that, a short discussion paper was produced and circulated, but with little positive response anticipated, given earlier experiences. PAGE 3 OF 54 However it clearly struck a chord in many places, because in fact, it generated unexpectedly enthusiastic support, endorsement of objectives, offers of cooperation and even funding. In this way, an international cooperative effort to tackle causation in complex systems was launched. A core organising group was developed, mechanisms for handling the funding from several sources were created, and a program of international engagement and work started to grow. Details, including the original discussion paper, the organising group, and other materials are provided in Annex B. While this project was initiated by AG14, it has now transcended AG14’s bounds both with respect to participation beyond the five nations of TTCP, and with respect to scope beyond the domain of purely defence-related research. However, AG14 strongly feels that this is a necessary path to take for the very reasons of the magnitude and generality of the issues. Although AG14 has now reached the end of its 3 year mandate, the program described here is still gathering momentum and it is too early to attempt a synthesis of what has been learned. Accordingly, the present document is better described as a Progress Report. We expect that the program will continue and produce its intended deliverables well beyond the lifecycle of AG14. However there is a valuable opportunity here for a successor TTCP entity to continue the strong shaping, motivating and contributing role that AG14 has played, and to steer its course so as to ensure defence-relevant outcomes. Report Overview The structure of this progress report falls into three sections. The first section, after this context-setting introduction, discusses the challenge of causation in complex systems, and introduces the notion of networked causality, and the difficulties that poses. Given these problems, a discussion of the ways in which limited models can be used for decision support in complex situations follows. The capabilities and limitations of existing approaches are then briefly reviewed, including reference to one particular approach based on dynamic networks, which is discussed in more detail in a supplementary report. PAGE 4 OF 54 In summary, having identified problems and needs, this section suggests that although high fidelity detailed predictive models of complex situations are unlikely to be achievable, useful decision support is still possible in such situations in the context of an appropriately adaptive decision methodology, with more modest, and therefore more achievable, but not yet fully achieved, goals for the representation and analysis of complex situations. The second section proposes a two-pronged strategy for addressing those goals, of generalising from existing domain-specific advances gleaned from a state-ofthe-art review, and generating new approaches from a CAS science perspective. A way forward for the state-of-the-art approach is discussed, and the section concludes with an exploratory discussion of some promising CAS ideas, and of some consequent research questions. The final section addresses implementation of this program. A brief account is given of workshops held and insights generated. The section concludes with a discussion of the final deliverables aimed for, and how they are to be achieved. Detailed program documents including a list of open questions, and an invitation to participate are provided in Annex B. PAGE 5 OF 54 “The complexities of cause and effect defy analysis” Douglas Adams: Dirk Gently’s Holistic Detective Agency Causation in Complex Systems Influencing complex systems towards desired outcomes and away from undesired ones requires some understanding of how they develop in time, and what the consequences of events may be, given the system’s properties, current state and history. Traditionally, one looks for causal relationships between events and their consequences, but how does causality operate in complex interconnected systems? In fact, the concept of causality has a long history in philosophy and science and remains controversial to this day even in quite simple scenarios, let alone in complex ones. However the agenda here is to take a very pragmatic approach to causality as the production and propagation of effects, with the ultimate aim of supporting effective interventions in real world complex systems. Therefore we have not sought to delve into the formal or epistemological nature of causality for its own sake, but to nevertheless remain open to the possibility that some insights from such deliberations may illuminate the difficult problem of analysing possible future trajectories of a complex system and the consequences of various actions or events in order to support better decisions about possible interventions. We begin by examining the sources of difficulties. Causal and Influence Networks What makes systems complex is the network of interdependencies between the elements of the system. This means that consequences of any event or property develop through many interacting pathways, and similarly, that if we want to know how a particular property or event came about, we will find that there were many cross-linked pathways that contributed to it. In such a system therefore, we cannot expect simple causality (one cause – one effect), or linear causal chains (each effect inexorably causing the next, like a Rube Goldberg machine), to hold in general. PAGE 6 OF 54 Yet models based on simple causality and linear causal chains are attractive because they seem tractable, unlike the confusing complexity of the real system. It is possible to make a complex system appear simpler by restricting the scope of attention to a particular pathway, but if the scope is widened to include other pathways, or if unexpected side-effects that have propagated through those pathways are linked back and suddenly manifested within the restricted scope, we are quickly reminded that the causal chain was just one of many pathways through a network. Such systems are therefore inherently characterised not by linear causal chains, but by networks of causal relationships through which consequences propagate and interact. Such networksi of interactions between contributing factors can exhibit emergent behaviours which are not readily attributable or comprehensible. The simple observation that complex systems are characterised by networks of causal relationships has some significant consequences for our ability to understand and influence them: Properties of complex systems are the product of many causal pathways, which may themselves be cross-linked – thus influencing each other; Consequences of events or properties of complex systems may also develop through many causal pathways which may also be cross-linked and influence each other; Causal and influence relationships in a network can close into feedback loops that will amplify or suppress certain types of outcomes, and which can also interact with each other; Multiple interacting pathways exist in general between an event or property of a complex system at one time, and a property of interest of the system at some later time. i For example, the firing of a particular neuron in a brain is influenced by the input signals received from many other neurons, and its own firing in turn influences the probability of firing for many further downstream neurons, including some of those that provide its own inputs. Looking at the entire network of interconnected neurons we see firing patterns which are difficult to disentangle yet seem associated with complex higher-level functions. PAGE 7 OF 54 We are using the provisional term Causal and Influence Network (C&IN) to describe such phenomena, but other terms have also been proposed: emergent causality, complex causality, networked causality and decentralised causality. Challenges to Understanding C&INs It is not just the fact that causality is networked that makes C&INs a challenge to understand. Another fundamental property of real complex systems is that they exist at multiple scales, raising questions about how causal interactions at one scale affect or are manifested in behaviour and properties at other scales. This includes both emergence, the appearance of complex structure and dynamics at larger scales as a result of smaller-scale phenomena, and the converse, top-down causation, whereby events or properties at a larger scale can alter what is happening at the smaller scales. Traditional scientific reductionism partially addresses emergence by explanatory decomposition of complex phenomena to smaller scales, but for many reasonsii cannot predict what the emergent properties will be, given the known properties at the smaller scales. Top-down causation would seem essential to our ultimate objective of more effective intervention and influence of complex systems, but is in fact philosophically contentious, with some even denying its existence. Furthermore, there is no unique perspective which provides a complete view of a complex system, there is no unique way of determining its boundary, and wherever the boundary is drawn, interactions between the system and its environment are important. And in real complex systems, many properties and events will be inaccessible to observation, so one has to expect to always be working with uncertainty and incomplete information. All these factors add to the difficulty of studying the C&INs operating in a given complex system or situation. ii Reasons include the unpredictability of external and contingent factors that may influence the emergent outcomes, non-linearities that result in extreme sensitivity to initial conditions (therefore creating large uncertainties in computer predictions), incomplete knowledge of the C&IN, and in many cases, practical uncomputability in reasonable time. PAGE 8 OF 54 By contrast, much of our cultural conditioning is predicated on a naïve view of linear causal chains. Examples include: finding ‘the cause’ of an effect, or ‘the person’ to be held responsible for something, or ‘the cure’ for a problem etc. This focus on singular or primary causes makes it more difficult to intervene effectively in complex systems and produce desired outcomes without attendant undesired ones (so-called ‘side-effects’ or unintended consequences). Yet effective ways of interacting with particular complex systems have been developed in many domains, such as medicine, personal relationships, politics, business, genomics, agriculture, etc. While these domains are certainly not free of unintended consequences, neither are people paralysed into inaction by the bewildering difficulty of understanding the complex systems they are interacting with. On the other hand, our limited ability to understand and purposefully influence situations and systems that have large scale complexity hampers our effectiveness in dealing with them – witness for example the attempts to contain the recent cascading financial crises, ongoing problems with terrorism and insurgencies, difficulty in motivating concerted international action on climate change, world poverty, and many other serious global challenges. Similarly, in the defence sector, the problems of creating, transforming, integrating, managing and employing our extensively networked forces also stem from their underlying complex C&IN structure. Our hypothesis is that there is a need not only to challenge inappropriately linear causal thinking in complex systems, but importantly, to search for and develop where necessary, better ways of thinking about causation and influence in complex systems, in order to help achieve more effective action in these problem situations. The many issues that need to be confronted in attempting to rise to this challenge can be broken down into four key areas: 1. Methodology for engagement with complex C&INs: categorisation of causal and influence networks and of our relationships with them – i.e. understanding what kind of problem we are dealing with, and hence ... understanding what is and is not possible to do, and hence … developing appropriate meta-strategies for engagement. PAGE 9 OF 54 2. Learning about the C&IN: choosing system boundaries, finding sources of information, choosing what to observe, and the resolution of observations, use of probing actions, developing and testing conjectures about the C&IN, determining what can be inferred and predicted, identifying critical uncertainties, identifying and learning about important interactions through the system boundaries . 3. Modelling the C&IN: representing information and conjectures about the C&IN, dealing with information, dealing with multiple perspectives at multiple scales, and in particular, the fact that the overall C&IN in a complex situation often consists of many coupled C&INs which are also multi-scale and diverse (eg economic aspects, social aspects, infrastructure aspects etc) finding useful levels of abstraction, working with heterogeneous approaches, representing the C&IN’s environment (which is also a C&IN in its own right), exploiting and interrogating the model, and identifying what further information is needed, developing and testing techniques for reasoning about and exploring the possible behaviour and properties of a C&IN, negative, uncertain, contradictory and incomplete 4. Interacting with and Influencing the C&IN: developing and evolving measures of success and failure for intervening in the system, developing and assessing options for stratagems to influence the system, developing proxies for success and failure associated with the stratagems, PAGE 10 OF 54 implementing and evolving a stratagem, determining what needs to be monitored to support continuing evolution of the stratagem and of the underpinning conjectures. In each of these main areas and their subsidiary issues, advances are needed at the theoretical and conceptual levels, as well as practical tools and techniques to aid practitioners. The goal of this project is to make some progress towards assembling a generalised set of approaches, concepts and tools, and to stimulate further research and development to continue improving the toolset available for supporting complex system practitioners. Decision Support in Complex Situations To discuss the practical feasibility and potential utility of tools for supporting decisions in complex situations, it is necessary to first make a distinction between two classes of models, which we might in this context call system models and situation models. System models (SMs) represent well-defined systemsiii, with clear input and output parameters, based on physical models of the processes that transform inputs to outputs. A user can therefore run the model with certain input parameters and 'discover' the resulting output parameters, which could in principle be compared to real world outcomes and hence are falsifiable, provided the input parameters used were an adequate representation of the real world system and relevant aspects of its environment. On the other hand, a complex situation requiring decisions to be made, will comprise many interacting systems and be characterised by C&INs. A situation model to provide a complex decision support environment (CxDSE) must therefore contain many SMs and the interactions between them, as well as their interactions through the boundaries of the C&IN, since we know that such situations can never be treated as closed. In other words, a CxDSE is itself a virtual complex system. iii If the system in question is also a complex system, and if it is further characterised by a C&IN in its own right, then the boundaries become blurred and the complexities compounded, so let us reserve the term system model in this context for those cases where the description given here is adequate PAGE 11 OF 54 Designing and implementing a CxDSE is therefore partly an exercise in learning about and modelling a C&IN as discussed above, and a far more difficult challengeiv than developing the component SMs. The issues involved include all those raised in the preceding section methodology (deciding how to create the CxDS/E), framing (what SMs should be included? How should they be represented? What interactions should be included, both between SMs and with entities outside of the frame?), and so on, plus the overall integration issue – how to put the CxDSE together in a way that achieves fitness for purpose. Since the complexity of this task mirrors to some extent the complexity of the decision situation, and since system-of-systems integration challenges have repeatedly defeated us at much lesser levels of complexity, the difficulty of the integration issue should not be underestimated. However, study of the evolution of biological complexity imply that such integration challenges can be handled by iterative adaptive processes where the 'fitness' of the emerging complex system is continually tested and the feedback used to guide incremental integration and build. So if this is an effective methodology for developing the CxDSE, then it also informs what is required of the component SMs to achieve overall fitness for purpose. It is in the context of such an approach that the uses and utility of models can now be discussed. A falsifiable SM that is developed to closely match the real world system it represents can be used with high confidence in both the qualitative and quantitative aspects of its outputs. However the more SMs are interconnected to represent complex situations, the more rapidly uncertainties accumulate and multiply. Moreover in reality, many of the component systems of the overall situation will themselves be complex systems which need to be treated as C&INs in their own right. So a recursive approach also needs to be applied of using SMs to develop models of C&INs and then treating them as component systems to be integrated into situation models of the bigger context in which they are embedded, and so iv There is a deep-seated parallel here between the difficulty of complex systems engineering (building an integrated or federated system-of-systems from many diverse component systems), and that of building an integrated or federated CxDSE of C&INs from many diverse SMs. Similar approaches are called for and similar problems should be expected. PAGE 12 OF 54 on. The complexities and uncertainties are compounded at each step, so a CxDSEv for a real situation will almost inevitably provide low (on the quantitative aspects) to mid (on the qualitative aspects) confidence at best, even with high confidence in all component SMs, and even lower confidence if components systems are complex rather than SMs. So there are two related issues here: how to develop CxDSEs to be as good as they can be, and how to make the best possible use of the limited CxDSEs that can be built. Answers to the second question have consequences for what counts as fitness for CxDSEs, which in turn has consequences for the fitness required of the component systems and of their integration. Figure 1 illustrates how a range of models may be used. A high confidence model can be used with low risk to support big costly (to implement and to reverse) decisions, provided one had high confidence that the model was correctly framed for the decision. Revisit model for option generation better options Invest in reducing sources of critical uncertainty low Confidence entertainment mid Types of outcomes high Identify possible +/‘side-effects’ & long-term effects Invest in ability to look out for them Exploit /nurture or Suppress /mitigate identifying where further investigation is needed Narrow range by getting feedback as you go Many incremental reversible & modifiable Low risk support to big costly decisions Ranges of outcomes Decisions All subject to C/R/B big costly (to implement & to reverse) Figure 1 Uses of Models in Decision Support. See text for details. v Such an environment could in fact support many different roles or objectives: educating users about the possibilities in the complex situation, exploring the intrinsic dynamics of the modelled complex situation, providing a simulated environment to support experimentation with, learning about, and practice in, metacognitive skills, methodology, strategy or tactics for dealing with real world situations with broadly similar characteristics, event reconstruction eg for accidents, or most challengingly, in fact supporting real decisions in the real world. PAGE 13 OF 54 At the other end of the scale, a low confidence model cannot be used for decision support, but could be used for other purposes such as entertainment, or to learn about the inner dynamics of the model itself (without extrapolating to conclusions about the real world). Many models however could fall in the middle range of confidence where one would be taking too much risk to base big costly decisions on their quantitative outputs. However they might offer sufficient confidence in their qualitative outputs i.e. their indication of the kinds of outcomes that could result from a decision, and the possible ranges of values of output measures, to still have some utility in decision support. Specifically, if a big costly decision is unavoidable, then such results at least highlight where further investigation may be required to reduce risk. If however a more adaptive stance can be taken by shifting the decision strategy away from big costly decisions to many incremental and modifiable decisions (subject of course to a rigorous Risk/Cost/Benefit analysis) then the model results can be used iteratively to narrow the range of uncertainty and to identify the possible side-effects and longer-term effects that need to be monitored for and that can be adaptively managed if and when they are observed. This is important because as we have argued, high confidence models of complex situations and C&INs are unlikely to be feasible. Furthermore, to the extent that the best that can be done really fails to meet minimal requirements for decision support, this can also stimulate further research to reduce some of the sources of uncertainty in the CxDSE. Another option, is to go back to the CxDSE and ask whether it can help generate alternate decision options that have lower associated risk and adequate benefits. Finally, and not shown in the diagram, an incremental and adaptive approach to influencing the complex situation can (indeed must) also be used to adaptively refine the CxDSE itself through comparison of actual outcomes with those predicted by the CxDSE on every incremental iteration. These different uses of models imply very different ways in which their fitness for purpose should be evaluated, and therefore different objectives one would strive for in their development. In summary, it is possible to conceive of CxDSEs which include a variety of SMs interacting in C&INs, and which can provide at best very limited confidence in PAGE 14 OF 54 the real world accuracy of the consequences they produce for various decision scenarios, yet which can be of great utility in supporting an adaptive methodology for understanding and influencing the complex situation. Existing C&IN Approaches – capabilities and limitations A number of well-known approaches are obvious candidates for dealing with C&INs. While they all offer some useful capabilities, they also have significant limitations with respect to the challenges discussed above. Bayesian techniques Bayesian [3] approaches can represent the combination of many feed-forward causal pathways, and can be quantified, provided subjective assessments of all the relevant priors can be made. In principle at least, such an approach can be placed in a real-time learning loop and the observed versus predicted posteriors can inform adaptive updating of the priors. However the requirement for an acyclic graph structure is restrictive since this means that feedback and more complex closed loop structures cannot be incorporated within the Bayesian model itself. This reflects the fact that the technique is really aimed at supporting inferences from uncertain data, rather than making forward projections about complex temporal dynamics. A detailed review [4] of commercial and government-off-the-shelf Bayesian and similar decision support tools was performed against a set of essential and desirable performance measures (listed in Annex A) for dealing with complex military operations. None was found that met all the essential criteria. Nevertheless this class of tools does offer valuable capability for interpreting what is known about a situation and highlighting inconsistencies between beliefs and the data. Applications could be developed that facilitate its real-time adaptation as the situation evolves. However it is not suitable for exploring complex temporal dynamics or for doing what-if analyses in C&INs containing adaptive agents. System Dynamics System dynamics [5], originally developed in the 1950s, is a widely used approach, amenable to quantification and simulation, and able to provide good visualistion of the dynamics of flows, their accumulation into stocks, with feedbacks and time delays, in C&INs. It can be used to explore different regimes of of system behaviour, provided one has accurate enough information about the PAGE 15 OF 54 relationships in the system, but cannot incorporate innovative adaptive processes which change the structure of the C&IN. Another drawback is that despite the ability to execute System Dynamics models to discover the outcomes for a given set of initial conditions, all this really does is trace one possible short path out of a vast possibility space. There are no systematic tools that enable analysis of the possibility space. Some modified versions of System Dynamics have been used to aid comprehension of complex operations with some success, such as TNO’s Marvel [6] Network theory A note of disambiguation is needed first. Although the word ‘network’ appears both in network theory and in AG14’s concept of C&INs, there is an important difference. In a visualisation of a C&IN the ‘nodes’ do not represent objects but properties or events which generate consequences on other properties or events, represented by the ‘links’ to other nodes. There is therefore an explicit temporal dimension along each path in a C&IN. What is networked is not entities, but the flow of causation and influence over time. Network theory [7] on the other hand, is the study of graphs as the representation of asymmetric relations between discrete objects. Many network analysis tools operate on time-integrated rather than temporally resolved network data (apart from looking at how different network growth rules result in different topologies) and they do not address emergent non-topological properties, such as innovations by adaptive agents leading to changes of structure and behaviours. Network tools have become very popular over the last decade since the discovery of small world and scale free networks, but they generally focus primarily on system properties arising from the topology of the network, rather than the functionality of nodes. Network theory is therefore well suited to addressing such aspects as social networks, communication patterns, traffic flows, propagation effects and network optimisation problems, but has little to say about how information is processed and used to drive purposeful change. One approach that does offer a richer and more dynamical view of networks is Dynamic Network Analysis [8], and a supplementary AG14 report [9] on application of this approach to defence problems accompanies this report. PAGE 16 OF 54 Simulation In practice, because of the lack of good analytical models of real complex systems, practitioners have long resorted to simulation, in which the complexity of the system or situation of interest is re-created dynamically, albeit in limited and approximate ways, rather than attempting to build it into a model upfront. Some of the approaches mentioned above lend themselves to simulation (such as Systems Dynamics and some network tools), however one pre-eminently CASrelevant simulation approach is the use of Multi Agent Simulation, or Agent Based Simulation, [10] applying the principles of autonomy, local views and interactions, and decentralisation. These approaches offer the benefit of resolving spatio-temporal patterns, but at the expense of stripping agents down to very basic features. For example, there is generally no multi-scale structure in an Agent Based Simulation, and their range of behaviours are very constrained. Nevertheless complex emergent behaviours such as self-organisation, which do not depend on detailed properties of the agents, can be produced. Some very large scale Agent Based Simulations are currently being developed to study the dynamics of real world C&INs, and they may well produce valuable insights if a very careful methodological approach is taken. However one drawback suffered by all simulations is that they essentially represent single trajectories through the phase space of the system rather than any systemic analysis of the phase space. A derivative of this approach, Agent Based Distillations [11] in which supercomputers are used to generate multidimensional data arrays about many parallel trajectories, were pioneered by the US Marine Corps Combat Development Command in the late 90s and early 2000’s, but currently still fall far short in many respects of addressing the C&IN challenges identified here. Some of the best examples of effective use of simulations to deal with real world C&INs are those developed for the real time adaptive management of epidemics, such as that developed at Argonne National Laboratory [12]. Non-Linear Dynamical Systems The theory of NLDS, [13] based on the notion of time evolution of a dynamical system through its phase space via a deterministic or stochastic rule, generates very powerful mathematical tools permitting analysis of the system’s trajectories, and recognition of attractors, bifurcations and the onset of chaos. However it PAGE 17 OF 54 cannot really be used for systems such as most real world C&INs lacking a mathematical formulation. In spite of not being able to exploit the rich mathematics of NLDS many of the concepts such as state space, phase space, attractors, trajectories and chaos, can be usefully applied in a loosely analogous way to real complex systems that cannot be mathematically represented. Some of these will be discussed in the section exploring CAS concepts. Discussion It is evident that no one of these approaches is able to address all the challenges discussed above, and some of the challenges are not addressed by any of them [4]. Admittedly, and as discussed above, there are nevertheless useful roles for most of them within the context of an appropriate CAS based methodology. However a critical gap remains in relation to innovative adaptation, complex temporal dynamics, and complex multi-scale emergent phenomena, that is unlikely to be filled by any combination of these approaches. New approaches are needed to understanding and working with C&INs, that are grounded in the science of complex systems and that can bridge the gap to access the many strengths of established approaches and augment them with these essential features of complex adaptive systems. Strategy There are two potential avenues for discovering better ways to understand and work with C&INs: I. State-of-the-Art review to leverage domain-specific advances already made in particular disciplines, and II. Novel CAS approaches – i.e. push the boundaries of the state of the art by exploring relevant ideas in CAS science that could lead to innovative advances. State-of-the-Art review This involves selecting a number of disciplines for review, and documenting how their practitioners deal with the challenges of networked causality in their systems of interest. To render their insights more broadly applicable they need to be documented in a generalised domain-independent way. PAGE 18 OF 54 Furthermore, it is also necessary to understand what it is about their systems of interest that make their specific approaches and techniques feasible and useful, so that the transferability of their approaches to other systems can be assessed. The implies a four-step methodology: A. Explore and review the state of the art of theoretical concepts and approaches, and practical tools and techniques (referred to as methods for brevity in what follows) in use across a number of disciplines that deal with complex systems. In particular, document how they: discover and model the C&INs operating in their systems of interest; explain how observed properties and behaviour have come about; think and reason about the possible consequences of observed or hypothetical states or processes; estimate the probablilities of various outcomes, and explore how hypothetical properties or events might come about. B. Generalise the identified methods, determine under what conditions they are valid, and hence the characteristics required of a C&IN for those methods to be applicable to them. C. Hence develop a classification of both C&INs and the generalised methods, according to those characteristics, so that a mapping can be made between them, to identify which methods are suitable for which systems. D. Discuss how to instantiate the generalised methods into applicable domains, and discuss how they enable the user to: define or recognise systemic (i.e. networked) causality in contrast to traditional notions of direct causality; assess the kinds of causal and influence relationships that can be identified; and capture, represent and investigate the C&INs operating in it. In order to help focus effort on areas of either or both greatest need and greatest potential for immediate benefits, we add the following two steps to the strategy: E. Develop a prioritised list of open questions in relation to C&INs in the disciplines consulted, and hence develop lists of both fundamental C&IN research questions, and domain-specific high pay-off C&IN questions. PAGE 19 OF 54 F. Look for cross-disciplinary opportunities to address any of the specific questions with what has been learned from steps A-D. Since the number of disciplines that could potentially be reviewed is very largevi, some selection criteria are needed to determine which ones in particular will be explored. The criteria identified and agreed are: Availability of SME’s in the discipline that can be consulted for the review, including both established practitioners and recent graduates; An assesment of how much there may be to learn from the field Whether the field offers coverage of generic problems, application issues or spatio-temporal scales not addressed by the other fields selected. Novel CAS Approaches This avenue of the overall strategy needs to be informed by the following considerations: Greatest needs: Appreciation of the priority fundamental C&IN research questions and high pay-off problems to be addressed; Current limitations of existing methods; and Promising CAS concepts that might be developed into useful innovations. The results of Step E above will provide the first of these, the areas of greatest need. The second consideration was discussed in the preceding section, and the third is now developed in more detail in the following. Promising CAS Concepts There are many complex situations where experts have developed effective ways of perceiving the essential aspects of those situations so that they can rapidly make good decisions about them. So an important question to ask is in what ways are experts who successfully deal with particular types of C&INs different from novices who struggle with them? vi Examples include Neuroscience, Medicine, Genomics, Defence, National Security, Government Policy, Ecosystems, Economy, Finance, Diplomacy, International Relations, Parenting, Education, Psychology, Business, Infrastructure management, Development and Aid Relief, etc PAGE 20 OF 54 The evidence suggests that experts make sense of complex phenomena in their domains by developing higher-level conceptual structures that reduce the hyperdimensionality of searches for solution or strategy elements. For example a chess grand master does not see a chess board in play as just the sum of the positions of the individual pieces, but as coherent larger scale patterns which can be reasoned about, whereas the novice who can only see the individual pieces is overwhelmed by the combinatorial explosions of possibilities. Experienced doctors similarly can quickly diagnose complex conditions from patterns created by a few observations whereas the recent graduate has to exhaustively investigate all the possibilities with tests and questions. By extrapolation then, a reasonable goal might be to discover a set of CAS-based mental constructs which aid in rapid appreciation of the structure of a large amount of interconnected information about a C&IN and clarify its essential dynamics and possibility space. So if this is a successful way of making complexity more manageable, what are the relevant higher-level concepts for C&INs? CAS science suggests that interactions and dynamics in C&INs are more important than composition and static configurations. But the main sources of dynamically organised complexity are adaptive, self-organising and regulatory processes, so could these be the building blocks? and if so, how might they be used to develop better conceptual structures to deal with networked causality? To explore this idea further it is necessary to understand how such processes interact, and what recurring and/or persistent patterns and structures they can create. Can we develop a “physics of interacting adaptive and S-O processes”? There are two promising options that have started to be explored. However this section is still purely speculative and needs development and experimentation. Multi-scale Intent This concept is discussed in AG14’s companion report on Complex Operations [14]. Its relevance here is that it may make a number of simplifications possible. For example, sets of agents with similar multi-scale intent structures at a given scale of description may be aggregated and treated as a distributed super-agent, reducing the number of entities to be considered. PAGE 21 OF 54 The combination of an agent A who at one scale can effectively control the finer scale intents of other agents a1, a2, … (whether by coercion, reward or inspiration) can be replaced with a single agent with the intent of A and the enlarged capabilities of the set. When combined with the adaptive processes whereby agents’ intents are changed, the dynamics of a complex situation with a large number of agents might be replaced with a much smaller number of variable composite interacting agents at a several scales. Attractors Attractors are regions of the possibility space of the situation which are more likely to be occupied than other regions, and in the extreme, once entered, are very difficult to escape from. In Non Linear Dynamical Systems theory, attractors are well-defined mathematical constructs which characterise the system’s asymptotic phase space trajectories as the net result of all the forces acting. Such attractors are therefore manifestations of equilibrium seeking physical processes. Examples include point attractors, limit cycles and the Lorenz strange attractors. However in complex adaptive systems, agent behaviours are driven by intents and opportunities not simply by physical forces (although they are of course bound by the laws of physics, their behaviours are not comprehensible from that perspective alone). We claim that adaptive processes of agents (whether individual or collective) create “attractors” by virtue of the fact that they exert a bias for particular outcomes and against other outcomes, so distorting what might otherwise have been an equi-probable range of outcomes. The quotation marks are to acknowledge that this is a qualitatively different phenomenon than the mathematical attractors mentioned above, although there is an obvious analogy. In the formal mathematical sense, what creates the attractor are the restoring equilibrium seeking physical forces that bring the system back when a perturbation causes an excursion. In complex adaptive systems, what creates the attractor are the dynamic adaptive processes of agents acting in their own interests. This creates a dynamic stability as opposed to an intrinsic lowestenergy stability. PAGE 22 OF 54 The attractor created by a single agent might be quite weak, but if many agents’ adaptive processes are coherently aligned then the collective result may be quite a deep attractor. However such a coherent alignment would be vulnerable to sudden collapse if a changed condition affected them all in exactly the same way. A more robust attractor might result from the interactions of a number of different but interlocked adaptive processes whereby a perturbation that threatened to weaken any one element would stimulate adaptive responses from those others most immediately affected, which in turn would stimulate others, so that successive synergistic waves of adaptive responses would overcome the perturbation and restore, perhaps even deepen the attractor. Such attractors would be manifested as very persistent features of the C&IN which one would describe either as resilient if they were perceived as desirable, or intractable, if they were seen as problems. Many examples spring to mind, e.g. the intractability of corruption, drug use, poverty and crime problems. Each of these can be thought of as an attractor and its persistence understood as resulting from the interacting adaptive processes that are stimulated by any attempt to modify the dynamics. Interestingly, many of the individual agents whose adaptive behaviours contribute to the dynamic stability of the attractor may in fact be trapped by it themselves. It is also evident that such attractors can interact at a larger scale and create more systemic attractors (e.g. corruption protecting and profiting from crime which protects and profits from drug use, which contributes to and exploits poverty which feeds corruption etc) This is just one example, but the principle worth exploring here is that there may be a number of robust features of C&INs which result from the myriad ways in which adaptive and self-organising processes could interact in mutually stabilising ways. Discovering these might be a key to developing those higher level constructs which could replace the overwhelming and confusing diversity and multiplicity of individual entities and events with a much richer but simpler picture of the essential dynamic interactions of fewer significant elements overlaid on a background of static that can be filtered out. PAGE 23 OF 54 The underlying fundamental research challenge here is to develop a general theoretical framework for understanding the interactions between adaptive and self-organising processes. What is known about interacting adaptive processes? Specific examples of interacting adaptive processes, and adaptive processes interacting with self-organising processes, have been examined in a number of fields. For example: Population dynamics study the co-evolution of species, in other words how two (or more) evolutionary processes interact. Arms races and the Red Queen effectvii are both examples of competitive evolution between two species, and can also be seen in competitive adaptations in two agents (e.g. a literal arms race) Group selection in evolutionary theory (the interaction between adaptive selection processes occurring at the individual level within a species, and at the group level within a group-competitive environment) has recently been rehabilitated by DS Wilson, Sober [15, 16], and others after years of mainstream rejection. New analyses of these interacting adaptive processes at different scales offer instructive examples. In fact multi-level selection theories [17] are now a hot topic in theoretical biology so there is a lot of new thinking that can be exploited. Niche Construction is the term given to the relatively new theory that explicitly addresses interaction between genetic evolution and the adaptive processes that shape the environment inherited by progeny, thereby modifying the selection processes operating on them. Extensive mathematical analyses of these very different interacting adaptive processes result in counterintuitive and very significant effects on evolutionary outcomes [18]. Even more ambitious studies of interactions between four different inheritance systems (Genetics, Epigenetics, Behavioural, Symbolic) are now being looked at [19]. The Red Queen effect is named for the character in Lewis Carroll’s Through the Looking Glass, who proclaims that “it takes all the running you can do, to keep in the same place”. Clearly the evolutionary consequernce of an arms race. vii PAGE 24 OF 54 The AG14 Conceptual Framework for Adaptation [2] recognises five nested levels of adaptation of which 4 involve applying adaptation to adaptive processes. These represent a subset of the ways in which adaptive processes can interact - a carefully selected subset that can operate simultaneously and synergistically. So-called “evo-devo” studies address interaction between evolutionarily developed self-organising processes and clade selection processes in development [20]. There is a rich and rapidly developing literature here that could and should be mined for important insights into how to formulate a more general theory of interacting adaptive processes. Such a theory could yield a wealth of novel and valuable ideas that could be applied to the problem of understanding and influencing complex systems. This is probably the single most important theoretical research challenge that has come out of this thrust, and advances could have immediate applicability in better ways to disable undesirable attractors in complex operations including use of IEDs, government corruption, cultivation of narcotics and so on, and helping to chart better strategies to foster the emergence of more acceptable attractors in counterinsurgency and interventions in failed states. It could also have applicability to many other domains, including organisational change, policy development, and complex systems engineering. Particular issues range from understanding the consequences (especially unintended) of interventions designed to address climate change, to improving the effectiveness of aid agencies, to name but a few. Implementation: The C&IN Program This ambitious program is clearly beyond the capacity of AG14 to implement alone. However, AG14 has been able to seed interest in a much larger community which is now providing funding, support and active participative contributions. Scoping Paper and Open Questions A short scoping paper circulated in early 2009 stimulated positive responses from many quarters, resulting in over $100K of funding being raised from the US Air Force Research Lab’s Asian Office of Aerospace R&D (AFRL AOARD), the US Office of Naval Research – Global (ONR-G), the Australian Commonwealth PAGE 25 OF 54 Scientific and Industrial Research Organisation (CSIRO), the University of South Australia (UniSA), the Australian Research Council (ARC, via the Complex Open Systems Network (COSNet) which they fund) and the Defence Science and Technology Organisation (DSTO). To promote discussion, a prioritised list of open questions relating to C&IN in relevant disciplines was developed. This was useful in identifying the very broad range of issues which people felt were relevant, and in stimulating interest in attempting to answer them. The scoping paper and the open questions are listed as part of the Invitation to Participate in Annex B. Workshop Series The funds raised were primarily intended to support a workshop series whereby leading international scientists can be engaged to participate and contribute. An Organising Committee was set up, co-chaired by the AG14 Chair, a senior CSIRO scientist and two US professors. The OC has arranged four workshops so far, listed below. The workshops have been reported in AG14 reports as AG14 activities. Table 1: C&IN Workshops held up till May 2010 Workshop 1. 2. 3. 4. 23-25 February 2009 First, inaugural, workshop on causal and influence networks, held under the auspices of Complex 2009 in Shanghai, China Mt Tamborine, Qld, Australia July 2009 Focused 5 day workshop with 17 participants from AS, US, CA and France Arlington, VA, USA Held in association with the AAAI November Symposium - Complex Adaptive Systems and the Threshold Effect. See [21] Shanghai, China 5-7 November 2009 Paris, France November 2009 Held at the Institut des Systèmes Complexes Paris in association with the International Conference on Emergence in Geographic Space: Concepts, Methods and Models PAGE 26 OF 54 There will be at least two more workshops in the second half of 2010. The purposes of the workshops are to build and develop a broad community of scientists and thinkers to engage in an extended dialogue about the questions raised here and in the Open Questions, growing collective understanding, and sharpening the research questions. Accordingly the workshops have been structured to foster critical debate with participants presenting brief summaries of relevant papers and short thoughtpieces around new ideas or questions to stimulate longer discussion sessions. For more information see: http://cs.calstatela.edu/wiki/index.php Insights from Workshops so far A lot of ground has been covered and intriguing ideas put forward. To give a sense of the breadth, a smattering of dot points highlights follow, but there is no attempt yet to form a coherent view. Prof Cliff Hooker (Uni of Newcastle, NSW, Australia) – bio-cognition, adaptive management of CS. Prof Sandra Mitchell (Uni of Pittsburgh, US) – causation and explanation in biological systems. Prof George Ellis (Uni of Capetown, South Africa) – 5 kinds of top-down causation in CS, theory of intervention. Aaron Bramson (Uni of Michigan) – mathematical formulation & analysis of tipping points. Dr Beth Fulton (CSIRO, Australia) – adaptive modelling and participative/adaptive management of socio-economic-ecological systems. Dr Anil Seth (Uni of Sussex, UK) – Granger causality in neuroscience. Nicholas Brodu (Uni of Rennes, France) – conceptual frameworks for causation. Dr Valeriy Vyatkin (Uni of Auckland, NZ) – Causality Analysis Using Model-Checking. PAGE 27 OF 54 Dr René Doursat (CNRS, Ecole Polytechnique and Director, Complex Systemes Institute (ISCIF), Paris, France) – Causing & influencing patterns by designing the agents. Prof Jacques Dubucs (Uni of Paris I – Sorbonne, Head Institute of History and Philosophy of Science and Technology) – Unfolding Complexity (knowing simple mechanisms in CS is not sufficient). Prof Hernan Makse (City College of New York, US) – Modularity map of C&IN (human embryo development). One strong theme did coalesce in the 4th workshop – the central role of the human mind and the multiple ways in which it was relevant to the program objectives: 1. Human Mind as a C&IN i.e. as an object of study, probably the most complex C&IN we know of, the consequence being that neuroscience, psychology and cognitive science are all fertile domains to consult 2. Human Mind as Decision-Maker i.e. as the target of the study, since it is human minds that we wish to aid in doing the understanding and decision-making to exert influence, the consequence being that is necessary to understand their strengths and limitations in order to deliver more useful understanding and decision aids – this links in directly with the AG14 thrust on Complex Cognition and Decision-Making [23] 3. Human Minds within a C&IN i.e. as elements of C&INs of interest, it is generally human behaviour in C&INs that add greatly to their intransigence, the consequence being that it is likely to be important and valuable to develop complex adaptive and multi-scalar conceptual models of human agents in C&INs (as discussed in [24]), and 4. Human Mind as Researcher i.e. as the subject of the study, via the minds of the study participants. This will be used as an organising and integrating theme in the development of the final products. Deliverables and Forward Workplan This project began with a very practical goal – improving the ability to understand and influence complex situations. PAGE 28 OF 54 The journey of seeking insights and novel ideas that might contribute to that goal has meandered through a very wide landscape of human knowledge, experience and reflection. At present (May 2010) we stand somewhere close to the widest perspective, which will be progressively focused as the key ideas and questions emerge. The closing phase of the project over the next 1-2 years, will be a process of abstraction, generalisation and synthesis. In the closing session of the 4th workshop in Paris, the group revisited its objectives and planned deliverables and outlined a forward workplan. A summary is presented here. The main outcome, apart from the individual learnings of all the participants, will be a publication of one or two volumes of peer reviewed and edited invited and contributed papers that provide a useful guide to best practice, priority challenges, and stimulate further research and exploration of novel approaches. The draft structure of this publication is in five sections: 1. Introduction: a. Statement of the Problem, b. Scope and Issues, c. History (incl of this project) 2. The State of the Art a. A number of compact subsections, one per selected knowledge domain, in which the following are discussed: i. Types of C&INs ii. Methodologies, tools and techniques iii. Why they work for these C&INs (and what the limits are, to what extent do they not work?) iv. What the high pay-off problems are v. Optional: illustrative examples and case studies provided by domain experts could be included in sidebars in some cases PAGE 29 OF 54 b. Synthesis: i. Approaches, tools and techniques identified in the review will be summarised and generalised, together with discussion of the conditions under which they are valid; ii. These results will inform a mapping between: o The generalised tools and approaches etc, and o Classification of types of C&INs iii. Summary discussion of domain-specific and domain independent (common) high pay-off issues 3. Human Factors in Interacting with Complex systems a. Discussion of how this is an integrating / organising theme for the whole project b. How the strengths and limitations of human decisionmaking should impact the way in which the results of this project are presented and developed 4. Novel Approaches It is hoped that a number of novel approaches will be proposed in addition to the speculative ideas put forward in this progress report. This section will consist of a number of invited and contributed papers and thoughtpieces. 5. Synthesis and Future Directions PAGE 30 OF 54 References 1. TTCP JSA AG14 Final Report, May 2010 JSA-AG14-AR-2010 2. TTCP JSA AG14 Synthesis Report: Conceptual Framework for Adaptation, May 2010 DOC – AG14 - CFA #3-2010 3. Nicholson A and Korb K Bayesian Artificial Intelligence Chapman & Hall 2004 4. Gossink, D et al Assessment of Causal and Influence Network Modelling tools relevant to OPSTSR 69 DSTO Internal Report [CLASSIFIED] 2007 5. Sterman J System Dynamics modelling: Tools for learning in a complex world California Management Review, 43 (4) pp8-25 2001 6. Van Zijderveld, EJA Marvel – principles of a method for semi-qualitative system behaviour and policy analysis, TNO, NL www.tno.nl/downloads/def_alg_Paper_MARVEL_SDS_20071.pdf 7. Newman M, Barabasi A-L, Watts DJ The Structure and dynamics of Networks Princeton University Press 2006 8. Carley, KM Dynamic Network Analysis, NRC Workshop on Social Network Modelling and Analysis (R. Breiger and KM Carley Eds) 9. TTCP JSA AG14 Supplementary Report, Pestov, I. CAS in Defence and Security: a Network-Enabled Approach, May 2010. DOC-AG14-C&IN:NEA #5-2010 10. Epstein, JM. Agent Based Computational Models and Generative Social Science. Complexity, Vol 4, # 5, pp 41-60, 1999 11. Horne, G Beyond Point Estimates: Operational Synthesis and Data Farming Maneuver Warfare Science 2001 USMC Project Albert, Quantico, VA 12. ANL Episim ref 13. Katok, AB and Hasselblatt, B Introduction to the Modern Theory of Dynamical Systems, Cambridge University Press, 1999 14. TTCP JSA AG14 Synthesis Report: Complex Adaptive Systems Concepts for Complex Operations, May 2010 DOC – AG14 - CxOps #2-2010 PAGE 31 OF 54 15. Sober, E. and Wilson, D.S.: 1998, Unto Others – the Evolution and Psychology of Unselfish Behavior, Harvard University Press, Cambridge. 16. Wilson, D.S.: 1990, Weak Altruism, Strong Group Selection, Oikos 59, 135–140. 17. Okasha, S Multi-Level Selection, Co-variance and Contextual Analysis Brit. J. Phil. Sci. 55 (2004) 481-504 18. Odling-Smee, FJ, Laland, K and Feldman, MW Niche Construction: The Neglected Process in Evolution Monographs in Population Biology, Simon A. Levin and Henry S. Horn (Eds), Princeton. 19. Jablonska, E and Lamb, M. Evolution in four Dimensions: Genetic, Epigenetic, Behavioral and Symbolic variation in the History of Life Life and Mind: Philosophical Issue sin Biology and Psychology, MIT Press, 2006 20. Kirschner, M.W. and Gerhart, J.C., 2005. The Plausibility of Life – Resolving Darwin’s Dilemma. Yale University Press, New Haven and London. 21. AAAI (2009). Complex Adaptive Systems and the Threshold Effect: Views from the Natural and Social Sciences. In Proceedings of the AAAI Fall Symposium: Arlington 2009, Mirsad Hadzikadic and Ted Carmichael Eds, AAAI/MIT press. 22. Mitleton-Kelly, E. The Emergence of Final Cause in Aaltonen, M., (Ed) The Third Lens. Multi-Ontology Sense-Making and Strategic Decision-making, Ashgate Publishing Limited, UK 2007 23. TTCP JSA AG14 Progress Report: Complex Cognition and Decision-Making, May 2010 DOC – AG14 - CxDM #6-2010 24. TTCP JSA AG14 Synthesis Report: Complex Adaptive System Concepts for Complex Operations, May 2010 DOC – AG14 - CxOps #2-2010 Additional Reading Dowe, Phil. Causal Processes, The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), Edward N. Zalta (ed.). Buchanan, B.G. and Shortliffe, E.H, Rule-based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Addison-Wesley, reading, MA, 1984. PAGE 32 OF 54 de Kleer, J. & Williams, B.C., Diagnosing Multiple Faults, Artif. Intel. 32, 97 1987. Hitchcock, Christopher. Probabilistic Causation, The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), Edward N. Zalta (ed.). Kaminska-Labbe, R., On the Co-Evolution of Causality: A Study or Aristotelian and other Entangled Influences, Acad. Management, Annual Meeting, Atlanta, 2006 Malinas, Gary and John Bigelow, Simpson's Paradox, The Stanford Encyclopedia of Philosophy (Winter 2008 Edition), Edward N. Zalta (ed.). Mancosu, Paolo, Explanation in Mathematics, The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), Edward N. Zalta (ed.). Mayes, G. Randolph, Theories of Explanation, The Internet Encyclopedia of Philosophy, J. Fieser (ed.). Mitchell, S., Exporting Causal Knowledge in Evolutionary and Developmental Biology, Philosophy of Science, 75 (December 2008) pp. 697–706. Nordin, Ingemar Complex Causation and the Virtue of Pluralism Medicine, Health Care and Philosophy, 9:321-323, 2006 Pearl, Judea, Causality, see http://bayes.cs.ucla.edu/BOOK-2K Reiter, R., A Theory of Diagnosis from First Principles, Artif. Intel. 32, 57-95, 1987. Rittel, Horst and Melvin Webber, Dilemmas in a General Theory of Planning, Policy Sciences 4 (1973), 155-169, Elsevier Scientific Publishing Company, Amsterdam. Schaffer, Jonathan, The Metaphysics of Causation, The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), Edward N. Zalta (ed.). Shalizi, Cosma R. Bayes < Darwin-Wallace, Blog entry. Wagner, Andreas, Causality in Complex Systems, http://www.bioc.uzh.ch/wagner/papers/BiologyPhilosophy1999.pdf Woodward, James, Causation and Manipulability, The Stanford Encyclopedia of Philosophy (Winter 2008 Edition), Edward N. Zalta (ed.). Woodward, James, Scientific Explanation, The Stanford Encyclopedia of Philosophy (Spring 2009 Edition), Edward N. Zalta (ed.). PAGE 33 OF 54 Annex A: Essential and Desirable Requirements for a C&IN Support Tool for Complex Operations Feature How this feature supports the Adaptivity Battle Urgency What tool should do Desired output What users can expect to able to do Show causal network on standard laptop screen. Causal graph that visually depicts the relationships between nodes Some customisation. Drag node locations, hide or show nodes. Users can read the text from within the software Users can edit and add text. Present data visually on a single screen and print. Visual representation aids comprehension. Single screen summary view aids easy briefing for CO and others. High Include narrative text Allows someone who was not involved in creating the model to understand it. Tells a story that is consistent with the causal graph. High Implement version control of models, including metadata. Adaptation requires change. An audit trail on the history of the causal network allows the user to understand changes and the progress of the model. High Allow users to access previous or alternative versions of the model. This could be sophisticated (eg CVS) or rely on users to develop naming conventions and use Save As to start new branches and significant updates. A set of model files that can be loaded one at a time. Each model has metadata so it can be uniquely identified User decides when significant changes merit a new version, and enters metadata. Show difference between two versions. Better helps understanding in previous item on version control. Very Low Allows a user to see what has changed between two versions. List of differences between two models. Two versions to compare. View temporal dynamics of variables Time scales and time lags are critical in understanding feedback. High Graph variables over time. A temporal view of individual variables. Select variable to view graphed over time, and set time window i.e. add annotations Export to a printable format. Allow entry of free text. Text can be associated with the whole model, or with components of the model. PAGE 34 OF 54 Identify and display feedback loops Feedbacks create nonlinearity. They require careful attention so that points of leverage and stability can be identified to focus adaptive action. Medium Feedback loops are shown in a list or highlighted in the visual display. The kind of feeback (+/-) would be also shown. List of classified feedback loops, with attributes . Or drawn as a layer on graph. Run algorithm to ID loops. Choose whether to hide or show loops in visualisation Identify and display internal inconsistencies; If the model is inconsistent, it may lead to invalid inferences about what action to take. Low If the user enters a set of hypotheses that are logically impossible, the tool should flag this, and require re-entry. Error message describes inconsistent model components User decides if error or if alternate models are needed. Display strength and nature of relationships Aids comprehension. Low Provide visual representation of strength of relationships, so that stronger relationships are bolder and +/- are different colours Variable width arcs in causal graph Enter strength Represent confidence; User knows relative confidence to place in different models, based on evidence Low Provide a confidence level for the whole model, and for components of the model. Confidence in [0,1] Enter evidence Show causal networks at various scales, allowing aggregation, hierarchical views and zooming Aids comprehension. Structure may be obscured by hundreds of variables. Give a sense of structure and importance Low Like Google Earth, some links and nodes can only be seen as the user zooms in. They are hidden to give a synoptic view. Each node has a scale attribute, and is hidden at coarser scales. Partial visualisation Either choose a scale, which determines automatically what is displayed; or manually show/hide nodes Facilitate search, query, and ability to select a variable and highlight every node within x Increases efficiency of user who wants to understand or change the model. Low User inputs a query which changes their view of the data. Can be used to hide information or focus on a region of the causal network. Partial visualisation Query that determines the focus and scale of observation PAGE 35 OF 54 hops Link visual components to narrative components Switching between visuals and text aids user synthesise information for deeper comprehension. Very Low Double clicking on a node or relationship pops up explanatory text, documenting the source and reason for the node/relationship Text boxes that can be shown/hidden within the visualisation. User enters free text. Support multiple concurrent users with different views of the same causal network Enables adaptation to occur at multiple levels and in parallel. Low Manage simultaneous read/write requests from multiple users. Messages such as: File currently locked by user X. Open for read only? Whether they require read or write access on Open Represent logical relations between hypotheses, eg AND, OR, NAND, NOR and XOR Enables more sophisticated relationships to be expressed. Low Allow input of coupling between relationships. Visualise couplings between relationships User specifies the couplings Display spatial information in GIS, and animate temporal data. Aids comprehension. Shows spatial and temporal patterns that may otherwise be missed Very Low Include a GIS mapping layer, or export data in a GIS format. Run temporal dynamics with movie style functionality (play, pause, step) Map based animated visualisation Each variable and relationship must have associated space and time attributes input Allows multiple concurrent users with competing theories that may involve inconsistent views of the same causal network Can evaluate the performance of multiple models from the same evidence. Inconsistencies between models informs important information requirements Very Low Allows users to create a new variant of the model with some different nodes and relationships. New evidence can update multiple concurrent models. Multiple models. Comparison between models produces a table of inconsistencies. User creates more than one model. PAGE 36 OF 54 DESIRABLE CHARACTERISTICS Feature Intent Importance User Documentation Is documentation available? If so is it in an accessible format and what degree of specialisation is assumed? High Training Documentation Is there a training program for the tool that is well established and known to provide users sufficient knowledge and expertise to use the tools effectively. High Need for Technical Understanding Does the tool require the user to have a substantial understanding of the mathematical techniques being employed and the implications of these Techniques. Medium Intuitive Interaction Are the menus, tool selectors, dialogue boxes etc, in places that a user may expect them High Convenience Can a user import extant models, are there clipboard like features eg. Cut, Paste, Undo and Redo. Medium Extensibility / API’s Is the tool under consideration easily extended or supplied with a set of API’s so that it can be modified to suit end application. (Who owns the tool and how accessible is it.) Medium Models instantly recognisable and meaningful. Are the models understandable without substantial interpretation? High Platformdependence Is the tool only available on certain platforms and does it function identically across platforms. Low Objective Correctness Has the tool been shown to be objectively correct? i.e. under a controlled environment provides confirmed correct results. To what degree are scientifically validated techniques utilised. High Evidence Inclusion Does the tool explicitly allow a user to introduce empirical evidence back into the model High Distributed Can the tool support multiple users simultaneously Low Part of a Tool Suite Is the tool self contained or part of a series of tools with extensions (eg. BPE and Matlab). Low Scalability Is the tool capable of handling large models? Medium Computability Can the tool compute on large models, does it scale linearly, or nonlinearly. Medium PAGE 37 OF 54 Annex B: International Collaborative Research Program Membership of Organising Committee Co-Chairs Anne-Marie Grisogono David Batten Prof Russ Abbott Prof Paul Davies DSTO Australia CSIRO Australia California State University, LA, USA Beyond Center, ASU, Phoenix, AZ, USA Members Axel Bender Berryman Mathew Bob Bolia Bohdan Durnota David Sonntag Fabio Boschetti Jimmie McEver Hussein Abbass Patrick Beautement Scott Wheeler Vanja Radenovic DSTO Australia UniSA Australia ONR USA Tjurunga Research Pty Ltd, China AFRL USA CSIRO Australia EBR, USA UNSW Australia Abaci Partners UK DSTO Australia DSTO Australia Invitation to Participate On the following pages is the letter and information that has been sent to the scientists so far invited to participate. PAGE 38 OF 54 International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks 8 March 2016 Dear colleague I am writing to invite your participation in an international collaborative research program on Causality in Complex Systems over the next 12 months. As you will be well aware, we currently face many serious global problems whose underlying dynamics are better described as Complex Causal and Influence Networks than as linear causal chains. Addressing such problems calls for better ways of learning about, understanding and influencing situations where causal pathways are so interconnected that traditional planning and management approaches fail. With the generous support of the research agencies shown below, and under the auspices of The Technical Cooperation Program Joint Systems Analysis Action Group on Complex Adaptive Systems for Defence, we aim to involve a number of leading researchers in relevant fields and to stimulate lively and deep exchange of ideas between them through a series of workshops and through online interaction. The overall intention is to review and advance the state of the art in understanding and influencing complex systems, culminating in an edited collection of invited and contributed papers. Your understanding of complex system dynamics would be very valuable, and we would really appreciate your perspectives on our research questions. Realising that there are many demands on your time, we are very flexible about how you might engage further with us, and are happy to discuss ways to facilitate your interactions with us. More information is provided on the next page and in appended documents, including scope, details about the planned workshops and other ways of participating. As a first step we would appreciate an indication of whether you wish to accept this invitation, and if so, what form that might take. Looking forward to hearing from you. Respectfully Anne-Marie Grisogono On behalf of the Causality in Complex Systems Organising Group Prof Hussein Abbass, UNSW Prof Russ Abbott, US CalStateLA David Batten, CSIRO (co-chair) Patrick Beautement, UK, Abaci Matthew Berryman, UniSA Axel Bender, DSTO Bob Bolia, US ONR Global Fabio Boschetti , CSIRO Marie Grisogono, DSTO (co-chair) David Sonntag, US AFRL AOARD DSTO Alex Ryan, SAMS, US Army DSTO AnneVanja Radenovic, Scott Wheeler, International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks Program Scope To initiate the program, we have developed a short scoping paper and a set of Open Questions (both attached). These documents, together with other information, are also available on the program wiki: http://cs.calstatela.edu/wiki/index.php/Causality_in_complex_sytems We invite you to share your thoughts in response to whichever questions attract your interest by email to the Co-chairs below. We expect the questions to evolve through the course of the program, so we also welcome your suggested revisions and new questions. Workshop Series 4 workshops have already been held, in Shanghai (as a special session of Complex’09, Feb 09), in Mt Tamborine, Qld, Australia 6 -10 July 2009, in Arlington, VA (as a session within AAAI Fall Symposium on CAS) 5 - 7 Nov 2009, and in Paris, France (as a session within Emergence in Geographic Space) 23-25 Nov 2009. At least two more workshops are planned, in Australia and the US, to progress and bring the program to a closing synthesis. Intended Outcome An edited 1 or 2 volume collection of invited papers addressing a coherent research agenda, evolved from our Open Questions, to be published by a major scientific publisher, in late 2011 or early 2012. We anticipate that participants will both learn from each other and generate new insights through the course of this program. The papers are therefore expected to be written after the workshops are completed, so that they include the new understandings that are reached. Ways of Participating Any combination of these is possible. Please feel free to discuss your preferences with us. A. Written response to some of the Open Questions in the attached document. B. Suggest papers to read and other contributors to invite. C. Participate in one or more of the workshops that are planned. D. Contribute a thought-piece to stimulate discussion during the program. E. Participate in online discussion through the wiki. F. Write a paper for the final collection of papers to be published. Papers may be contributed or invited. Invited papers will be sought from funded participants. G. Join the Organising Group, and / or become an Editor or Reviewer. Dr David Batten (Co-Chair) Dr Anne-Marie Grisogono (Co-Chair) Senior Research Fellow, Commonwealth Scientific and Industrial Research Organisation and The Temaplan Group T +613 92394420 M +61431708361 Research Leader Complex Adaptive Systems DSTO Long Range Research Fellow Chair TTCP JSA AG14 Complex Adaptive Systems for Defence Defence Science and Technology Organisation T +618 82596532 M +614 09076684 David.Batten@csiro.au Anne-Marie.Grisogono@dsto.defence.gov.au International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks SCOPING WHITE PAPER Research Theme on Networked Causality: Understanding and Influencing Systems with Complex Causal and Influence Networks Background: The concept of causality has a long history in philosophy and science and remains controversial to this day. This project does not seek to delve into the formal or epistemological nature of causality for its own sake, but rather to take a very pragmatic approach to causality as the production and propagation of effects, with the aim of supporting the understanding of real world complex systems. Of particular interest is the problem of analysing possible future trajectories of a complex system and the consequences of various actions or events in order to support decisions about possible interventions. What makes systems complex is the network of interdependencies between the elements of the system. This means that consequences of any event or property develop through many interacting pathways, and similarly, that if we want to know how a particular property or event came about, we will find that there were many cross-linked pathways that contributed to it. In such a system therefore, we cannot expect simple causality (one cause – one effect), or linear causal chains (each effect inexorably causing the next, like a Rube Goldberg machine), to hold in general. Yet models based on simple causality and linear causal chains are attractive because they seem tractable, unlike the confusing complexity of the real system. It is possible to make a complex system appear simpler by restricting the scope of attention to a particular pathway, but if the scope is widened to include other pathways, or if unexpected side-effects that have propagated through those pathways are linked back and suddenly manifested within the restricted scope, we are quickly reminded that the causal chain was just one of many pathways through a network. Such systems are therefore inherently characterised not by linear causal chains, but by networks 8 of causal realtionships through which consequences propagate and interact. Such networks of interactions between contributing factors can exhibit emergent behaviours which are not readily attributable or comprehensible. This simple observation has some significant consequences for our ability to understand and influence complex systems: 8 - Properties of complex systems are the product of many causal pathways, which may themselves be cross-linked – thus influencing each other; - Consequences of events or properties of complex systems may also develop through many causal pathways which may also be cross-linked and influence each other; - Causal and influence relationships in a network can close into feedback loops that will amplify or suppress certain types of outcomes, and which can also interact with each other; For example, the firing of a particular neuron in a brain is influenced by the input signals received from many other neurons, and its own firing in turn influences the probability of firing for many further downstream neurons. Looking at the entire network of interconnected neurons we see firing patterns which are associated with complex higher-level functions. International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks - Multiple interacting pathways exist in general between an event or property of a complex system at one time, and a property of interest of the system at some later time. We are using the provisional term Causal and Influence Network (C&IN) to describe such phenomena, but other terms have also been proposed: emergent causality, complex causality, networked causality and decentralised causality. A fundamental property of real complex systems is that they exist at multiple scales, raising the question of how causality at the lower levels is manifested in behaviour and properties at the higher levels. In addition, there is no unique perspective which provides a complete view of a complex system, there is no unique way of determining its boundary, and wherever the boundary is drawn, interactions between the system and its environment are important. In real complex systems, many properties and events will be inaccessible to observation, so one has to expect to always be working with uncertainty and incomplete information. All these factors add to the difficulty of studying the C&IN operating in a given complex system. By contrast, much of our cultural conditioning is predicated on a naïve view of linear causal chains – examples include: finding ‘the cause’ of an effect, or ‘the person’ to be held responsible for something, or ‘the cure’ for a problem etc. This focus on singular or primary causes makes it more difficult to intervene effectively in complex systems and produce desired outcomes without attendant undesired ones (so-called ‘side-effects’ or unintended consequences). Yet effective ways of interacting with particular complex systems have been developed in many domains, such as medicine, personal relationships, politics, business, genomics, agriculture, etc. While these domains are certainly not free of unintended consequences, neither are people paralysed into inaction by the bewildering difficulty of understanding the complex systems they are interacting with. On the other hand, our limited ability to understand and purposefully influence situations and systems that have large scale complexity hampers our effectiveness in dealing with them – witness for example the attempts to contain the recent cascading economic crises, ongoing problems with terrorism and insurgencies, difficulty in motivating concerted international action on climate change, world poverty, and many other serious global challenges. On the defence sector, the problems of creating, transforming, integrating, managing and employing our extensively networked forces also stem from their underlying complex C&IN structure. In all these arenas, better ways of understanding and working with C&INs might improve effective action. Approach G. Explore and review the state of the art of theoretical concepts and approaches, and practical tools and techniques in use across a number of disciplines that deal with complex systems. In particular, we will pay attention to how they: discover and model the C&INs operating in their systems of interest; Explain how observed properties and behaviour have come about; Think and reason about the possible consequences of observed or hypothetical states or processes; Estimate the probablilities of various outcomes, and Explore how hypothetical properties or events might come about. H. Generalise the identified theoretical concepts and approaches, and tools and techniques, determine under what conditions they are valid, and discuss their implications for how to: International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks define systemic (i.e. networked) causality in contrast to traditional notions of direct causality; assess whether cause-effect relationships can be identified and analysed in systems of agents that are changing interactively over time; Identify, capture and model (where possible) the C&INs operating in several systems of interest; I. Hence develop a set of classifications for C&INs that can be used to assess which theoretical concepts and approaches, and tools and techniques are applicable. J. Develop a prioritised list of open questions in relation to C&INs in the disciplines consulted, and hence lists of both fundamental C&IN research questions, and domain-specific high pay-off C&IN questions. K. Look for cross-disciplinary opportunities to address any of the specific questions with concepts or techniques generalised from other domains. L. develop and explore innovative CAS approaches to tackling selected fundamental questions. More Detailed Problem Statement There are four main areas in relation to C&INs where the project seeks to leverage existing knowledge and expertise in relevant disciplines, and to make progress in assembling a generalised set of approaches, concepts and tools: 5. Methodology for engagement with complex C&INs: categorisation of causal and influence networks and of our relationships with them, understanding what is and is not possible to do in different cases, developing appropriate meta-strategies for engagement. 6. Learning about the C&IN: choosing system boundaries, finding sources of information, choosing what to observe, and the resolution of observations, use of probing actions, developing and testing conjectures about the C&IN, determining what can be inferred and predicted, identifying critical uncertainties, identifying and learning about important interactions through the system boundaries . 7. Modelling the C&IN: how to represent the information and the conjectures, dealing with negative, uncertain, contradictory and incomplete information, dealing with multiple perspectives at multiple scales, finding useful levels of abstraction, working with heterogeneous approaches, representing the C&IN’s environment (which is also a C&IN in its own right), International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks exploiting and interrogating the model, developing and testing techniques for reasoning about and exploring the possible behaviour and properties of a C&IN, requirements for support tools. 8. Interacting with and Influencing the C&IN: developing and evolving measures of success and failure for intervening in the system, developing and assessing options for stratagems to influence the system, developing proxies for success and failure associated with the stratagems, implementing and evolving a stratagem, determining what needs to be monitored to support continuing evolution of the stratagem and of the underpinning conjectures. Examples of Well-known Existing Approaches – strengths and weaknesses Bayesian techniques – can represent the combination of many causal pathways, can be quantified, but cannot incorporate feedback loops System Dynamics – provides good visualisation of causal and inflience networks, can be quantified to explore different regimes of behaviour, but cannot incorporate innovative adaptive processes Network theory – many practical analysis tools have been developed, but focuses primarily on properties arising from topology rather than functionality of nodes Nonlinear dynamical systems – very powerful mathematical tools permitting analysis of the system’s phase space, but not applicable to C&INs lacking a mathematical formulation Interesting Questions to consider clarifying the relationships between predictabilty, control, influence and determinism, contingency, chaos, noise and uncertainty explore such notions of causation as: the deliberate initiation or stimulation of a causal network to produce an outcome (i.e. attributing intent and causal power to the human element); the lynchpin in a causal network, i.e. an element whose removal or modification results in a significantly different outcome; the ‘pawn’ who plays a role in a causal network but is entirely dispensable or replaceable i.e. removing the ‘pawn’ does not significantly change the outcome. How should we select the subset of C&IN dimensions that provide useful phase space views? can we identify those aspects of a particular C&IN which have greater and lesser degrees of predictability? Is there a way of analysing the topology of a phase space of a C&IN to identify regions of greater and lesser stability? And regions of greater and lesser influencability? In the absence of a mathematical (nonlinear dynamical systems) formulation of a C&IN, what utility do the concepts of attractors, tipping points and thresholds have? International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks Can we map adaptive and self-organising processes in a C&IN and associate them with attractors’ or regions of greater stability / predictability? Can we develop a “physics” of interacting adaptive processes? Outcomes and Deliverables Publication of a book (1-2 volumes) of selected papers by Springer, Elsevier or Wiley, addressing the above themes. Provisional assessment of whether cause-and-effect relationships can be identified in nonlinear systems of agents that are changing interactively. Produce a list of fundamental C&IN research questions, and high pay-off specific C&IN questions in the disciplines consulted Cooperative development of some spin-off benefits to selected specific C&IN questions in the disciplines consulted International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks Some Open Questionsix about Understanding and Influencing Systems with Complex Causal and Influence Networks (C&INs) These questions illustrate the tortuous journey that the organising group has been on in the last few months in formulating the issues that have brought us together. Fortunately, we are neither seeking consensus nor a definitive articulation of the problem space. We are however interested in stimulating some lively discussion and evolving a better set of questions, and in engaging wider and more diverse participation. We expect to learn many things as a result. They will all be interesting. We hope that some of them will also be useful to people who have to make decisions in complex situations, in spite of all the inherent pitfalls and uncertainties! Please respond to whichever questions motivate you. There is no expectation that anyone should respond to all of them. Answering even one is appreciated. If you work with real complex systems we are particularly interested in your responses to Part 2 (on the last page) as well. We will use your responses to evolve and focus our objectives and to plan the workshops throughout 2009/2010. We expect to have some funding to support invited participation in the workshops. Please support your responses where possible with references, and suggestions of other people who could further contribute. PART 1 - How does causality work in complex systems?? These questions are grouped into the following domains: Phenomenology – types of causation, heuristics 1. Are there different types of causation? Preamble: Different examples of causation that we easily recognise include: a deliberate intervention into a causal network to produce an outcome (eg a doctor prescribing a treatment for an illness); the lynchpin in a causal network, (i.e. an element whose removal or modification results in a significantly different outcome); the pawn who plays a role in a causal network but is entirely dispensable or replaceable (i.e. removing the ‘pawn’ does not significantly change the outcome.) These examples suggest two aspects of causation that may be relevant to our puposes: intentionality in the first instance, and causal power (high for the lynchpin and low fro the pawn). We would like to acknowledge that the format of this document is inspired by the ‘Open Questions on the Origin of Life’ document circulated by Pier Luigi Luisi and Kepa Ruiz-Mirazo to solicit participation in their workshop of the same title in San Sebastian in May 2009. ix International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks Question: Do you think it is important or valuable to make such distinctions? Why? Are there other distinctions that also need to be made? Do you know of ways to facilitate recognition of different instances, for example in a graphical map of a complex causal and influence network? Philosophy – framing, relation to explanation, ontology, computability, contingency and determinism 2. What do we mean by causation? Preamble: There are conceptual links between the notions of causation, predictability, control, influence, determinism, contingency, explanation and emergence, but the use of these terms and how they relate to each other is not always clear or consistent in the literature. This lack of clarity can be an obstacle to resolving important questions and to making the necessary conceptual distinctions. Question: What do you think the fundamental concepts are in relation to causation in complex systems? How do you see their relationships? Where do you think points of confusion lie and how do you think they should be resolved? 3. Computer Modelling of Causation Preamble: It has been argued that processes with genuine causal power can not emerge within a system of formal logic, such as all traditional computer programs are. A similar view in the field of causality is shared by Rosen x and Hausmanxi, among others, and is also somehow implicit in Hume’s classic discussion xii. This has crucial implications for complex causality; first because much research in Complex System Science is carried out via computer simulation, and second because a closed, formal logic approach is implicit in the abstraction of complex systems as existing in pre-defined state spaces. Questions: To what extent can causation be studied via computer modelling? What kind of causation can be studied via computer modelling? Has any computer modelling approach had success in examining causation? 4. Contingency versus determinism in the development of a complex system’s trajectory. Preamble. complex systems are often seen in terms of two basic, opposite schemes, determinism or necessity, and contingency. Generally, the two principles work hand in the hand, as each “choice” made by contingency must then comply to natural laws and, in turn, further contingencies arise through a developing history. However, if we ask whether the development of a complex system follows an obligatory deterministic pathway (absolute determinism), or whether it is due to the vagaries of contingency, the two views clash. The question. How do you see the interplay of these two principles in the temporal development of a complex system? How should we combine them in an approach to understanding and influencing complex systems? x Rosen, R. Life Itself, a Comprehensive Inquiry into the Nature, Origin, and Fabrication of Life; Columbia University Press, 2001 xi Daniel M. Hausman, Causal Asymmetries (Cambridge Studies in Probability, Induction and Decision Theory), Cambridge University Press, 1998 xii David Owens, Causes and Coincidences, Cambridge University Press, 2002 International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks 5. Framing Causation Preamble: The laws of physics (famously) don't speak in terms of causes. The equations (famously) work either forwards or backwards. And certainly one wouldn't want to attribute a teleological "cause" to events that produced known results. Perhaps we shouldn't use a causal-based framework for discussing the world. Questions: Do you agree that causal-based framework is not the best way of discussing the world? Is there an alternative? What about the way causation is treated in computer science, where one speaks of states instead of causes? Scientific – relation to physical principles and laws, roles of equilibrium, noise, energy, chaos, entropy etc 6. Equilibrium, causation, end energy flows Preamble: Is a perspective that starts from a preference for stability and equilibrium in fundamental conflict with the way the world we live in works? If there were no energy flows, there would be no causes and no effects (because nothing would happen) Questions: How would one relate energy flows to causality? If one presumes that all systems of interest involve energy being provided to them from an external source, does it make sense to attempt to understand such systems in terms of equilibrium-based modes? How should causation be related to energy flows? 7. Role of chaos, noise and uncertainty in causation Preamble: The notion of causation suggests reduction in uncertainty about the outcome whereas chaos and noise suggest the opposite. In real complex systems chaotic, causal and influence relationships coexist with sources of noise and uncertainty. Question: Do you think these antithetical aspects of the dynamics of complex systems should be treated independently (e.g., as simply placing computable limits on predictability?) or can they be integrated in some way? Are there fundamental research issues to be addressed here? 8. To change or not to change our minds Preamble: Several leading psychologists, neurologists and cognitive scientists have suggested that an individual agent (or a collective) can select what types of behaviour it wishes to promote or what types of behaviour it wishes to forbid. The particular parenting model we experience as children has a strong influence on what we regard as good or bad. There is evidence that those of us with conservative leanings tend to adopt the strict father model and shape their reasoning using use direct causation, whereas those with more progressive ideas tend to adopt the nurturant parent model and think quite differently about causation (systemic causality). Ultimately, conservative and progressive impulses may be rooted in our evolutionary programming. The tendency to create these kinds of classifications, and to act at least partially on them, seems to be universal and hard-wired in our brains. If this is true, the brain’s neural censors govern its decisions about which classes of behaviour to enhance or suppress. Questions: Why do some people reason using direct causation while others invoke systemic causation? When we change our mind about something, does it change our brains? Does the mental process of “changing one’s mind” correspond to the physical process of switching International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks between attractors in our brain? Would a complex systems perspective, centred on selforganizing, attractor neural networks, be helpful in ordering our thinking about how internal psychic and neurodynamic processes lead us to reason and behave in one way or another? 9. Connectionist Models for Decision-Making Preamble: Decision-making can be defined as an outcome of mental (cognitive) processes leading to the selection of a course of action amongst several alternatives. Those processes have historically been modelled in many different ways including: as following from individual aspirations, needs and goals; as a continuous process shaped from context of environment; or as a logical rational selection mechanism. More recently, connectionist models have sought metaphors for the functions of the brain (for example models of neural networks). Questions: To what extent do you hold that individual actions are shaped by their decisionmaking processes? What are the beliefs that make smart people dumb. Is any decisionmaking framework useful in identifying, diagnosing, classifying or otherwise assisting in coming to grips with causality in complex systems? Complexity – emergence, phase spaces, attractors, tipping points, thresholds, multiscalarity, ‘black swans’ 10. Emergence and emergent properties. Preamble. Emergence is a notion with many complex sides. A popular view is that emergent properties are those novel properties that arise when parts or components assemble together (in space and or time) into a higher hierarchic order – novel in the sense that they are not present in the parts or components. Ryanxiii has argued that emergence is coupled to scope rather than scale. Shalizixiv defines emergence in terms of predictive efficiency. Abbottxv defines emergence in terms of levels of abstraction. All these approaches address the presence of emergent properties of a complex system, rather than the process by which they may have arisen (emerged). In other words, it seems more natural to use the concept of emergence to refer to a class of properties rather than to refer to a class of process. Why would that be the case? Are the various processes that produce emergent properties more fundamental with emergence being no more than a short-hand for ‘here be dragons’? On the other hand, some argue that emergence is a fundamental defining aspect of complex systems. Maxent approaches for example claim that the underlying driver is the creation of more efficient ways of dissipating free energy and that emergence is the inevitable result. The question. How important is the role of emergence in driving the dynamics of a developing complex system? What is the relationship between emergence and the processes that produce the emergent properties? how should emergence be xiii Ryan, A.J. ‘Emergence is coupled to scope, not level’, Complexity 13(2):67-77, 2007. Crutchfield, J.P.; Shalizi, C.R. ‘Thermodynamic depth of causal states: Objective complexity via minimal representations’, Phys Rev E 1999, 59, 275–283. xiv xv Abott, R. The Reductionist Blind Spot, Complexity (forthcoming) http://www3.interscience.wiley.com/journal/121685801/ International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks integrated into a causal view of a complex system? What evidence can you provide for your views? 11. Complex Systems as integrated or loosely coupled Preamble. Complex systems can be treated as entities or as collections of coupled but still distinguishable subsystems, each with its own recognizable dynamics and requirements for stability. Both views are useful but neither is sufficient. To the extent that the entity view applies, the developmental trajectory of the system can be discussed in terms of entity level properties, and to the extent that the coupled subsystem view works, it may be amenable to analysing subsystem behaviours in somewhat independent terms and treating their interactions explicitly. Wilsonxvi notes that most systems of interest—such as biological and social systems—are simultaneously both entities and groups and that a key aspect of complex systems is to find structures and mechanisms that enable entities to form persistent groups. The question. Do you see these as complementary, alternative or inadequate views of complex systems? What do you think is the proper way to apply the notions of interdependency versus subsystem independence, in the understanding of both the historical and future development of a complex system? Can a different understanding of the organization and stability of a given complex system lead to better predictability of its future development? 12. Ecology and individuality Preamble: A (human) ecosystem can be understood as a collection of interactions and interdependencies which arise from the nature of the individual agents that participate in it. Alternatively, that same ecosystem could be understood as a collective, cooperative entity, whose structure and dynamics obey principles of their own; thereby determining the kinds of agents that can arise and survive within the system (in the sense assumed by Darwinian competition and selection). Certainly, ecosystems, their agents and the set of interactions between agents co-evolve over time. Does ecology come from interacting individuality, or does the converse apply? The direction in which we should infer causation is unclear. Questions: What are—and how should we discover—the organizational principles that govern collective ecology and individuality, and the directions in which causal and influence factors or constraints flow between these two scales? In what sense is individuality a subordinate concept to ecology, and in what sense is ecological order the subordinate concept? {UPWARD versus DOWNWARD causation} 13. Attractors and Phase Spaces Preamble: A nonlinear dynamical system that can be formulated in terms of differential equations permits application of very powerful mathematical techniques such as Melnikov theory to analyse its dynamics in terms of trajectories and attractors in its phase space. Most real complex systems of interest are not able to be expressed in precise mathematical form. Nevertheless it is often possible to recognise analogous properties such as thresholds, xvi Sober, E., & Wilson, D. S. (1998). Unto Others: The Evolution and Psychology of Unselfish Behavior. Cambridge, MA: Harvard University Press., International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks tipping points, and regions of greater likelihood which we might call ‘attractors’. In particular, complex adaptive systems have by definition some inbuilt biases for or against particular types of outcomes – i.e. those which their adaptive processes select for or against – which we could therefore identify as ‘attractors’ in some sense. A similar argument could be made for complex systems with self-organising properties, to associate their self-organised structures with ‘attractors’ or regions of greater stability/predictability. Questions: In the absence of a mathematical formulation of a C&IN, what utility do the concepts of attractors, tipping points and thresholds have? To the extent that you do see some utility to these notions, how do you think one should select the subset of C&IN dimensions that provide useful phase space views? How might we identify those aspects of a particular C&IN which have greater and lesser degrees of predictability? Are there ways of analysing the topology of a phase space of a C&IN to identify regions of greater and lesser stability? And regions of greater and lesser influencability? Do you know of any examples of such ideas being applied in particular complex systems? When there are more than one adaptive or self-organising processes active in a complex system, how should the effects of the multiple simultaneous processes be combined? Do you think it may be possible to develop a “physics” of interacting adaptive processes? 14. What Role does Interconnection Topology Play? Preamble: There has been a great deal of work in the complex networks community (see Boccaletti attached for an excellent review) regarding the role of the interconnection topology in determining the synchronisation properties of complex networks. Extended work by Siljak in large scale systems theory in the 80's and 90's also presented strong evidence of the relationships between many areas of interconnected systems theory and their graphical representation. There is a strong argument to be made that many more properties of complex adaptive systems can be understood purely in terms of the graphical topology of these systems. Question: How much information about the causality, operation, interaction and function of a complex adaptive system can be obtained from understanding and analysing the graphical topological structure of a complex system? Are there limits to what the topology can tell us about such a system? i.e. are there properties of such systems not revealed or only partially revealed from the topology and if so what are they? Pragmatics – methodologies for influence, tools, techniques, guidelines, human factors, errors and accident. 15. Finding useful subsets of complex problems Preamble: Research into complex decision making suggests that it is not necessary to completely understand the complex causal network underlying a complex situation to be successful in influencing its development. In fact what potentially separates successful decision makers in complex causal systems from unsuccessful ones is the ability to develop a useful conceptual model of a subset of the causal network that is adequate for their decisionmaking tasks without overloading their cognitive faculties. Question: Do you agree with these premises? Do you know of evidence to support or contradict these views? How do you think one should go about selecting the subset of a system to understand? Is there an optimal ‘size’ of conceptual causal network model for an International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks individual to apprehend and use in decision-making? What factors do you think influence this? 16. Humans as part of a complex system Preamble: Problems involving human action and human decision making are of particular significance in many serious and complex global problems. Questions: Focussing on humans as part of a complex system (that is, not as external actors) does it make sense to attempt to discriminate the effects of human intervention/action from the effects of inherent system dynamics? Can we identify states of the system in which human intervention is more likely to be effective? Can we identify a relation between the timing for human intervention and its spreading over the system at different time and spatial scales? 17. Prodding for answers Preamble: Trying to understand complex causal systems from the outside can be difficult, especially if the aim is to understand the system dynamics. Often the only way to do this is to act within the system in some way and observe the consequences. However this is turn changes the nature of the system. Question: Is it possible to understand a complex causal system without any kind of intervention? Do you know of any successful intervention approaches that have been used to understand causal networks? How can one intervene in a system and still get scientifically solid data about the system? 18. Visualising complex systems Preamble: Research has shown that people have difficulty understanding feedback and time delays, which are recurrent features of complex causal systems. However studies have also shown that at least in some cases performance can be highly dependent on the way the problem is framed or the interface used to display the problem. Question: Are there any “best-practice” techniques or methods for framing or displaying a causal network so as to assist the decision making process? What room is there for delegating some decisions to computers in areas where human decision making performance is poor? How can interfaces be designed better to accelerate understanding complex causal networks? 19. The Causation of Errors / Analysis of Error Causation Preamble: The large number of hardware and software components in a computer system can be thought of as a complex system, in that a large number of heterogeneous parts are interacting to produce output. We can examine causality in this complex system by looking at computer faults, and the process of tracking down the causes of these faults. In some cases we may end up with just simple causation - for example an error in a single line of code may cause an unexpected output from software. Often, however, there is no single cause - it may be an interaction between software running on the computer, USB device drivers, and printer firmware that causes a printing error. It is this causality in a web of linked components that we call “complex causality” Questions: What frameworks, methodologies, and tools are used to detect and analyse faults? How well do they handle complex causality? If any of them do handle it well, how do they go about their analysis? International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks 20. The Causation of Accidents / Analysis of Accident Causation Preamble: There are a large number of methodologies and frameworks for trying to identify the causes of accidents in systems. Many of these fall into a human trap of looking for a single cause. Questions: To what extent do some of the more modern systems approaches capture the fact that there may be a complex web of causal factors leading to accidents? Can you list examples of accidents where multiple social & technical causes have been found (in particular those where a “causal web” has been identified”), and by what methodologies they were found? 21. Defending against Black Swansxvii Preamble: “Black swans” are events that break a previously observed pattern (“all swans are white”) in esssentially unpredictable ways, thus demonstrating the dangers of inductive reasoning. While ‘black swans’ are difficult or impossible to imagine occurring (one cannot know if a black swan exists until the first one is spotted), they can in some circumstances be imagined. For example in sandpile, one can certainly imagine large avalanches but while it's just hard to predict them, it's not impossible to prevent them. At every time-step one could do a global analysis of where the dangers lie and eliminate them. (Of course one could always level the sandpile completely at each time step. But presumably one wouldn't want that sort of stasis either.) The problem is that such a global analysis and risk reduction work is relatively expensive. Questions: Has there been any effort to study, quantify or bound the predictability and preventability of such events? What do you think is possible to do and how should we go about it? 22. Role of AI in dealing with C&INs Preamble: Dealing with real C&INs will require us to make use of artificial intelligence and autonomous systems to complement and extend human capabilities so the ability of such machines to make sense of the complexity in any particular domain becomes a critical issue if they are going to effectively team with us. Questions: What can we say about the kind of human-machine teaming necessary? What is the level of what is realistically achievable? Could human teaming with such machines actually impair our ability to perceive the complex world – eg through leading us (because of their own inherent limitations) to only see deterministic aspects? To the extent that this happens, how might we mitigate its impact? 23. Optimal Location of External Control and Influence Preamble: In many large complex adaptive systems it is necessary to apply external control or influence to achieve a given goal. Existing work in control and influence of networks typically assumes that every element/node of the system can be controlled independently. In many complex systems of interest, especially biological systems, it is not possible to control each element directly but rather externalities can be applied at a limited number of locations xvii Taleb, N., The Black Swan: The Impact of the Highly Improbable, New York: Random House 2007 International Collaborative Research Program: Understanding and Intervening in Systems characterised by Complex Causal and Influence Networks in order for the whole network to perform as required. Given the magnitude and distributed nature of many of these systems it is interesting to know how best to apply these external influences to achieve a required goal. Question: How do we characterise the optimal location and magnitude of external control and influence to ensure a given complex adaptive system will function as desired? Is it possible to characterise networks for which it is impossible to provide external control and influence to achieve a given aim? In the case of limited external influence or control how should the control be applied? (i.e. is it best to apply the entire control energy at a single location? or is it better to have a larger number of smaller control sites?) 24. FINAL QUESTIONS: Have we missed something that you think is important? What additional questions would you like to see posed? What are your thoughts about answering them? PART 2 - What can we learn from disciplines that deal with real C&INs? Preamble: Many disciplines deal with C&INs in the real world. Examples include medicine, politics, finance, meteorology, social sciences, business, oganisational science, agriculture, genomics, ecosystem management, cognitive science, etc. Given that practitioners in these fields have been dealing more or less successfully with such real-world complex C&INs for many years, they have obviously developed domain-specific approaches that work for them, probably by taking advantage of specific features of the C&INs of interest to them. It is reasonable to expect that some of those approaches and concepts would be transferable to analogous complex problems in other domains, but to do that one needs to understand why particular approaches work and under what conditions. Questions: In your discipline, what theoretical and conceptual approaches are used to: discover and model the C&INs operating in your systems of interest? explain how observed properties and behaviour have come about? think and reason about the possible consequences of observed or hypothetical states or processes? estimate the probablilities of various outcomes? and explore how hypothetical properties or events might come about? What is it about these approaches that enables them to make sense of the complex phenomena? What 'mechanisms' are employed for engagement and influence? What do you think the wider community may be able to learn from this discipline?