Integrated Cognition AAAI Technical Report FS-13-03 What’s between KISS and KIDS: A Keep It Knowledgeable (KIKS) Principle for Cognitive Agent Design Benjamin D. Nye Institute for Intelligent Systems University of Memphis, Memphis, TN 38111 benjamin.nye@gmail.com Abstract Axelrod and Hamilton (1981) proposed KISS, which stated that agents should use the minimal number of variables and mechanisms then add new ones if needed. On the converse, the KIDS principle states that modelers should include all variables and mechanisms that appear relevant and then remove ones that do not add to the quality of the model (Edmonds and Moss 2005). Neither of these approaches are completely satisfactory for cognitive agents, which must combine general cognitive mechanisms with other task-specific factors relevant to the situation being modeled. A cognitive parameter that might be irrelevant for one task might be pivotal for another. Instead, cognitive agents are commonly developed using cognitive architectures. Cognitive architectures are usually designed to implement specific theoretical models drawn from cognitive science and psychology experiments. The architecture can then be applied to model specific cognitive tasks by specifying the task environment, altering initial states (e.g., initial information, social network ties), or specifying context-specific mechanisms (e.g., production rules for actions). Compared to a minimal model (KISS) or a maximal model (KIDS), this approach has the advantage that the architecture evolves in tandem with the theoretical model it was based on, such as ACT-R, Soar, or CLARION (Anderson 1996; Laird 2008; Sun 2007). Even normative (e.g., not biologically-inspired) models, typically adhere to standard theoretical or pragmatic constraints for their design. However, this approach suffers significant drawbacks. First, keeping a cognitive architecture up to date with new research is difficult and time-consuming process. Modelers work hard to stay current with literature, but each year the number of published studies increases and journals have steadily diverged into silos with more specialized focuses (Silverman 2010). As a result, cognitive models run the risk of quietly falling out of sync with new findings. Second, practices for comparing different cognitive models see limited use. Cognitive mechanisms seldom have quantifiable units and relationships between parameters may only be known at a correlational level (Silverman et al. 2001). As a result, there are usually infinitely many ways to implement the same combination of mechanisms. Model docking, where the equivalency of two models is tested for some task, has been proposed to compare models but docking tech- The two common design principles for agent-based models, KISS (Keep It Simple, Stupid) and KIDS (Keep It Descriptive, Stupid) offer limited traction for developing cognitive agents, who typically have strong ties to research findings and established theories of cognition. A KIKS principle (Keep It Knowledgeable, Stupid) is proposed to capture the fact that cognitive agents are grounded in published research findings and theory, rather than simply selecting parameters in an adhoc way. In short, KIKS suggests that modelers should not focus on how many parameters, but should instead focus on choosing the right research papers and implement each of their key parameters and mechanisms. Based on this principle, a design process for creating cognitive agents based on cognitive models is proposed. This process is centered around steps that cognitive agent designers are already consider (e.g., literature search, validation, implementing a computational model). However, the KIKS process suggests two differences. First, KIKS calls for reporting explicit metadata on the empirical and theoretical relationships that an agent’s cognitive model is intended to capture. Each such relationship should be associated with a published paper that supports it. This metadata would serve a valuable purpose for comprehending, validating, and comparing the cognitive models used by different agents. Second, KIKS calls for validation tests to be specified before creating an agent’s cognitive model computationally. This process, known as test-driven design, can be used to monitor the adherence of a cognitive agent to its underlying knowledge base as it evolves through different versions. Implications, advantages, and limitations of the proposed process for KIKS are discussed. 1 Introduction Cognitive agents are powerful tools for running simulated experiments, forecasting political outcomes, powering intelligent tutoring systems, and driving agents in virtual worlds (Sun 2006). However, significant debate exists over the appropriate design practices for such models. In particular, selecting the parameters and mechanisms of an agent’s cognitive model remain more of an art than a science. The agent-based modeling community has proposed two design principles at opposite ends of a continuum: KISS (Keep It Simple, Stupid) and KIDS (Keep It Descriptive, Stupid). Copyright c 2013, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. 63 niques are not well-standardized and are seldom performed (Burton 2003). Third, it is difficult to rigorously validate which relationships that a particular cognitive model incorporates as it evolves over multiple iterations. While the underlying theories and empirical relationships are reported in publications, these mechanisms are spread across many papers and technical reports. The purpose of this paper is to propose a process for developing cognitive agents meant to overcome these limitations. Cognitive agents should “Keep it Knowledgeable, Stupid” (KIKS). Unlike other domains often modeled using agents (e.g., political science, macro-economics), cognitive agents benefit from the knowledge provided by a wealth of relevant, well-controlled studies. These represent an embarrassment of riches: so many experiments are conducted that cognitive agents cannot even use all of this knowledge. The KIKS principle changes the focus from selecting the set of parameters (e.g., KISS vs. KIDS) to the selecting a set published papers that contain the theories and empirical findings used to design an agent’s cognitive model. These relationships capture the real knowledge that underpins the cognitive agent. By explicitly coupling the design of a cognitive model to this knowledge, cognitive agent design, comparison, and validation might all be significantly improved. 2 or scenario being studied. During each phase, the following tasks are performed: 1. Select papers and other information sources 2. Identify key parameters and relationships for each paper 3. Record structured metadata on each paper’s relationships 4. Create validation tests that check for these relationships 5. Implement parameters and mechanisms into a computational model In the first phase, the general mechanisms of a cognitive model are derived from core cognitive theories and empirical findings drawn from a specific set of published papers. These provide the requirements for the cognitive architecture for the model. While the definition of a “key parameter” or relationship is fuzzy at the ontological level, most well-written papers clearly state the main effects, interactions, or relationships that were studied. These requirements should be formally represented as metadata and associated with their published source. Based on these requirements, tests should be specified first and then the cognitive architecture should be created so that those tests should be satisfied by cognitive agents built using this architecture. In the second phase, context and task-specific knowledge is added, such as training data or studies findings relevant only to a specific task. The order of this process mirrors that of the core knowledge, with relevant papers broken down into a set of task-specific requirements that can be used to specify tests. This information can then be used to set up the starting states for the cognitive agent. These include parameter values, model structure inferred from a training data set, mechanisms present in the agents’ environment, or cognitive mechanisms that are too specific to the task being studied to be included in a more general model. After both phases are complete, the completed cognitive agent can be subjected to the general cognitive and task-specific validation tests specified during the earlier steps. Different cognitive models place a different level of emphasis on these phases. Cognitive models built using architectures such as ACT-R place significant emphasis on accommodating empirically-derived human cognitive principles (Anderson 1996). By comparison, a reinforcement learning agent for machine vision might require very limited cognitive restrictions (e.g., only those for reinforcement learning), but might require an extensive set of training examples or a carefully specified reward function for the task. So then, the level of effort required for each step will often depend on the purpose of the system. In large part, model designers already perform many of these steps. Two primary differences of KIKS should be noted, however. First, this process emphasizes producing explicit metadata about the theoretical and empirical relationships that the model is intended to implement. This metadata should be publicly available to other researchers and used to develop validity tests. Very few cognitive architectures state a process for collecting and recording data on the literature underlying the model. To this author’s knowledge, only the PMFServ cognitive architecture has collected metadata specifically on its source publications, through its KIKS: Keep it Knowledgeable The KIKS principle is that the design of cognitive agents should be explicitly and formally tied to their underlying literature, rather than implicitly tied to it. The foundation of all cognitive agents rests on published scientific papers. Cognitive modelers recognize the importance of theoretical grounding and coherence, taking pains to report the basis of models in their publications. However, the process of moving from literature to a validated cognitive agent is often informal. Based on KIKS, a structured process is proposed here to break the development cycle of cognitive agents into distinct phases and steps. Figure 1: Cognitive Agent Design Process Based on KIKS This KIKS process is outlined in Figure 2. The numbers in this diagram indicate a suggested order a two-phase design and validation process. These two phases differ by the knowledge considered: the knowledge behind the core architecture and the knowledge required for some specific task 64 eters and mechanisms can be added at a later time based on additional literature. Compared to KISS or KIDS, KIKS also offers the opportunity to improve the research ecosystem for developing cognitive agents. The following sections consider three areas where KIKS could improve upon existing modeling practices. First, the implications for KIKS on the goals of cognitive agent modeling are briefly explored. Second, the value of metadata for published papers is discussed. Last, the benefits for validation and model comparison are noted. anthology of human behavioral models (Silverman et al. 2001). However, this anthology was a collection of short reports, rather than formal, machine-readable metadata. With that said, the KIKS approach is significantly inspired by PMFServ’s process for breaking seminal papers down into key effects and constructing cognitive mechanisms tied to that literature (Silverman et al. 2006). However, KIKS is intended to be architecture-neutral, which will be discussed more in later sections. The second major deviation from standard cognitive agent design is that the specification of the validation tests occurs before the computational implementation of the cognitive model or agent. This approach is inspired by test-driven design, a software design pattern where requirements and validation tests are built first, then the system is built to conform to those standards. While test-driven design is uncommon in software design, it is very well-suited for cognitive modeling. Fundamentally, a computational cognitive model is built based on two things: 1. relationships derived from well-established, peer-reviewed research and 2. the modeler’s best-guess at how to resolve ambiguities, gaps, and underspecified mechanisms between this research. Clearly, the former must be trusted far more than the latter. A testdriven design process prioritizes the established cognitive mechanisms as constraints for model development, representing how the model should work after it is designed. It also puts an up-front focus on what it means for the cognitive agent properly implement its underlying theory and relationships. Moreover, this process can be used to consider the implications of additional papers, which can be broken down into requirements and used to develop tests to determine if (or when) the agent’s cognitive model conforms to these findings. This is an explicit form of the scholarship often performed with cognitive theories, where their implications are compared against new empirical studies. Cognitive models, much like cognitive theories, do not need to agree with every study. However, it is very important to understand which findings agree or disagree with a given cognitive model, as these establish boundary conditions and drive scientific discussion. Finally, compared to KISS or KIDS, this process greatly simplifies parameter and mechanism selection. Under KIKS, a cognitive agent should employ the parameters presented in the publications used to develop the model. For each paper used to support the model, the designer should identify the key parameters studied, empirically-derived relationships between them, and functional mechanisms proposed by theory. These parameters and their relationships should be explicitly represented as metadata before starting development of the computational model. So then, instead of choosing parameters or mechanisms, the model designer is actually choosing published references and explicitly deriving the model from them. In this way, literature search directly drives model design. Other parameters or mechanisms might also be added to the model (e.g., for efficiency or taskspecific needs), so long as the model still behaves properly and fulfills its specified requirements. For example, as part of the two-phase development process, task-specific param- 3 The Objective of a Cognitive Agent The KIKS process also has some implications over the objective of cognitive agent design. Cognitive agents serve three main roles: an explicit representation of existing knowledge, a predictive tool to discover new knowledge (e.g., through simulation), and an artificial intelligence to power human-like applications (e.g., intelligent tutors, virtual agents). The first role, as a representation, acts as a constraint on design: the cognitive model should embody prior findings and well-established theorized relationships. The second role, as a predictive tool, acts as an objective of the design: to be able to forecast real-life outcomes. The third role, as an AI, acts as an alternative objective: to be able to emulate or replace a real human agent. Both of the later two roles are objectives tied to a particular task, such as generating the distribution of human reaction time on a digit recall experiment (a predictive objective) or passing a Turing test (an AI objective). In both cases, the objective is connected to the task for the cognitive agent. So then, it is possible to consider a cognitive agent in terms of an optimization problem where the cognitive model for an agent is selected to optimize or satisfice some taskspecific objective function (FT ), but is subject to the con−→ straints (CC ) provided by the research findings and theory that it was based on. The objective function is some measure of performance on one or more tasks, which establish the domain the cognitive agent is intended to handle. For example, the objective of a cognitive agent for digit recall tasks might be to minimize the difference between the distribution of recall times for the model versus an empirical distribution collected from human subjects. While an some individual cognitive models only focus on a single task, cognitive architectures are often intended for broader scope of tasks or even a unified theory of cognition (Newell 1994). −→ The constraints (CC ) are the cognitive requirements that fundamentally make an agent’s cognitive model represent the cognition of a human or other agent, as opposed to a general artificial intelligence for performing certain tasks. While the objective function asks, “Is the model getting it right,” the constraints check, “Does it get it right for the right reasons?” In theory, a purely normative model does not need any cognitive constraints: if only needs to perform well on its tasks. However, in practice, even machine learning models designed often use variants of biologicallyinspired mechanisms such as reinforcement learning (Kaelbling, Littman, and Moore 1996). Also, normative assumptions are also typically rooted in published theories (e.g., 65 rational actors). Overall, to design an effective cognitive model, some constraints on the model are almost always required to keep the system theoretically coherent and computationally tractable. A second set of constraints also exists −→ for the specific task (CT ). These constraints might be environmental rules or boundaries of the task, for example. The design of a cognitive agent selects the parameters of −→ the cognitive model (XC ), state transition functions (e.g., update rules, a transition matrix) that relate these parameters (PC ), and the distribution of initial states for the cognitive −→ model (SC ). Likewise, the task environment must be de−→ signed, with its own parameters (XT ), state transition func− → − → tions (PT ), and distribution of initial states (ST ). For the task state and cognitive state to interact, a subset of state transition functions must use parameters from both. Transitions that change cognitive state based on task state are often → − called observations ( O ), while those that consider cognitive state to change the task environment are often called actions → − ( A ). For this discussion, we will assume that a modeler primarily has control over the design of the cognitive model −→ −→ (i.e., XC , SC , PC ), which corresponds with the first phase of the KIKS process. With that said, the dynamics due to observations and actions are also sometimes considered part of the cognitive model (which would make them “free variables” for modeling also). Assume that the intention of the modeler is to minimize the difference between the cognitive agent task performance and some objective, as well as to sat−→ −→ isfy a set of boolean constraints (CC and CT ). This produces an optimization problem in the form shown in Equation 3. minimize − − → −→ XC ,SC ,PC subject to ter DataShop for educational data mining (Koedinger et al. 2011). Secondly, as an optimization problem, it is possible to consider the potential for constraint relaxations, such as variants of Lagrangian multipliers (Everett III 1963). By relaxing constraints (e.g., replacing them with weighted penalties on the objective function), the formulation supports quantifying the tradeoffs of adhering to specific literature findings versus performing optimally on a specific task. Moreover, not all theories or empirical findings are equally trusted or required in a system. As such, tying penalties to confidence in research findings and theories offers one way to capture these intuitions. Alternate representations, such as Bayesian formulations, could also be used to unify constraints and objectives into a weighted objective function. As a grounding example, assume a cognitive agent that wanders a maze looking for rewards with two cognitive constraints: novelty-seeking (orient attention toward less familiar stimuli) and motivated attention (orient toward rewards). Since these constraints would be based on empirical studies, they might be recorded such as “The correlation between attention and novelty was 0.7, with a 95% confidence interval of ±0.2” (within a particular study). One way to interpret this constraint would be to require the cognitive agent to maintain a correlation of 0.7±0.2 for at least 95% of runs through the maze, rejecting non-conforming agents as invalid. A similar constraint could be constructed for noveltyseeking. By adding more constraints, the space of valid cognitive agents shrinks and created agents should better approximate the subjects from the underlying literature. For certain agents, this would be an ideal approach (e.g., cognitive agents that represent rats, with constraints based on studies of rats in mazes). However, the all-or-nothing approach gives constraints effectively infinite importance: no valid agent can violate them. For some agents, such constraints may impose unsatisfactory impediments to task performance (e.g., gathering rewards). Alternatively, certain constraints may be less relevant to the cognitive agents’ contexts (e.g., findings from laboratory studies used to build a cognitive agent for more naturalistic settings). In these cases, relaxed constraints that penalize the objective function for violations might be more appropriate. The relative weight given to the original objective function versus the total weight for relaxed constraints would determine the importance of task performance versus fidelity to literature. Similarly, differential weights for constraints would capture a relative importance for satisfying each constraint. While more work is required to explore best-practices for building agents that directly incorporate findings from studies, formalizing the purpose of a cognitive agent implies that existing optimization and machine learning algorithms can be applied to the problem. Finally, since this formulation focuses entirely on representing and testing against published literature, it is applicable to any cognitive agent design. Two major dimensions of cognitive models for agents are their scope and their theoretical coherence. The scope of a cognitive model determines its generality. Some models are intended to be a universal model of human cognition (general), while others only −→ −→ −→ − → → − → − FT (XC , XT , SC , ST , O , A , PC , PT ) −→ −→ ai (XC , SC , PC ) = 0, ai ∈ CC −→ −→ −→ − → → − bj (XC , XT , SC , ST , O , → − A , PC , PT ) = 0, bj ∈ CT KIKS posits a cognitive agent design, regardless of its implementation details, is implicitly searching for a satisfactory solution to a problem in this form. In this view, failed validation tests correspond to constraint violations and model performance corresponds with the objective function. This implies two corollaries. First, it is theoretically possible to design agents that explicitly optimize over this problem. This means that an architecture could be developed that generates cognitive models based only on metadata about published papers and a training data set for modeling performance of certain tasks. While such a system would almost certainly overfit a single task, one that was trained on a variety of related domain tasks might produce a useful cognitive agent while reducing authoring effort. Such a system is one way to combine knowledge from well-controlled experiments with big data, which is becoming increasingly important due to social network data, crowd-sourced experiments using systems such as Mechanical Turk (Rand 2012), and initiatives such as Pittsburgh Science of Learning Cen- 66 model a single interaction from an experiment (specific). Framing KIKS as trying to optimize over an arbitrary group of tasks (T ) captures this concept of scope. Some systems, such as machine vision, treat this set of tasks very explicitly and have canonical databases for training and testing. In other systems, the scope of design may only be in the modeler’s head and recorded in their published works. Theoretical coherence refers to the neat versus scruffy continuum (Minsky 1991). Neat models, such as dynamical systems, rely on a small number of mechanisms whose implementation can often be tested using proofs. Scruffy models use a larger breadth of literature and mechanisms, whose implementation must usually be tested using validation tests on the completed system. Obviously, both dimensions lie on a continuum and hybrids are common, such as a neat, generic core cognitive model supplemented by scruffy, task-specific mechanisms. While neat and scruffy models may use different techniques to validate their mechanisms (i.e., implement the functions in CC and CT ) and measure their performance (FT ), any two models based on the same requirements (e.g., theory, empirical findings) that have the same scope of tasks (T ) are effectively trying to solve the same problem. As such, it should be possible to base either type of design on the same metadata that describes these requirements and objectives. This implies that the metadata that describes the specifications for a cognitive model may have significant value for the cognitive modeling community, as it could be shared by different architectures approaching similar problems. The importance of such metadata is described in the next section. 4 onomies for cognitive science are currently an active area of research. In neuroimaging (e.g., fMRI studies), the BrainMap taxonomy describes two decades of key components in published results and the more recent Cognitive Paradigm Ontology was developed to structure this into a domain ontology for describing neuroimaging experiments and results (Turner and Laird 2011). The Cognitive Atlas is a similar project that contains an ontology for cognitive science specifically, with over 400 tasks and 600 concepts currently represented (Poldrack et al. 2011). Finally, the Cognitive Modeling Repository is a recent project dedicated to storing implementations of cognitive models (Myung and Pitt 2010). While this project does not have an explicit ontology for parameters, it does require structured metadata reporting the input and output parameters of simple cognitive models. There is undoubtably overlap of the concepts represented in these projects with those used in the cognitive models used by cognitive agents. As such, they might provide an excellent starting ontology for metadata about key relationships presented in journals. Journals have also increasingly recognized the value of metadata. Structured abstracts are the most pervasive method for structured data about publications, with significant use in medicine, education, biochemistry, and other domains. A structured abstract breaks an abstract down into required fields (e.g., Background, Objectives, Methods, Results, and Conclusions). Structured abstracts not only improve the ease of selecting relevant articles (Nakayama et al. 2005), but there have been proposals for journals to hook these into digital summaries and ontologies to provide semantic search capabilities (Gerstein, Seringhaus, and Fields 2007). Some biology researchers have even directly textmined publications to find patterns, though this approach is likely limited to domains where the objects of study (e.g., protein sequences) have clear, unique names (Altman et al. 2008). With that said, the majority of biology metadata has been created by human curators and offers an example that could be imitated in cognitive modeling. KIKS would benefit from structured metadata for new journal publications, so moves by relevant journals in this direction would strengthen this approach. Reciprocally, KIKS would generate metadata about existing publications when they are used to design cognitive agents. Obviously, this short discussion only scratches the surface of the challenge of general metadata for cognitive research. A database of metadata on findings and theorized relationships would have long-term value for cognitive science and psychology as a whole. This section is primarily meant to show how KIKS could contribute to and benefit from such a repository. However, even in absence of a shared metadata format and repository, ad-hoc metadata that describes how a cognitive model connects with literature would be a valuable resource for understanding and validating a single project. These issues are addressed in the following section. Knowledge Base on Cognition The first three steps of a KIKS phase (Select Papers, Identify Relationships, and Record Metadata) focus entirely on distilling the knowledge present in publications. Since the findings from publications are constant, formal representations of these relationships could be shared by different cognitive architectures. If this metadata were stored in a shared repository, a structured knowledge base of cognitive science findings and theoretical principles would slowly emerge. The value of such a knowledge base cannot be understated. The value of even being able to search for papers based on their findings would be enormous and might eventually facilitate real-time meta-analysis capabilities. However, for such metadata to be useful, a workable domain ontology of cognitive concepts, experiments, and results would need to be available. Moreover, building this knowledge base does not necessarily need to be the modeler’s responsibility. Academic journals and authors of scientific papers should have an interest in making their findings easy queried and betterintegrated with other findings. Even putting aside cognitive agents, the argument can be made that this type of work is vital to managing the flood of scientific papers and findings published each year. For this to be sustainable over the long term, journals need to encourage structured data about published findings and theories. Over the last twenty years, steps have already been made in this direction by related fields. Ontologies and tax- 5 Validation and Comparison By specifying requirements before designing a cognitive agent, KIKS makes the validation process more principled. 67 KIKS suggests that authors should represent their underlying theoretical and empirically-derived relationships before implementing a cognitive model, then test the cognitive agent against these requirements. This allows for test-driven design, where tests are actually written before the implementation of the computational model. This approach can be done iteratively, where additional literature findings provide new constraints that the cognitive model should satisfy. Unlike the specific cognitive model implementation, these constraints do not need to be specified at the level required to create a computational model. Instead, they only need to be specified at a level that allows a test to validate that the constraint holds. For example, studies on motivated attention have found that being hungry correlates positively with additional attention to videos or pictures of food (RoskosEwoldsen and Fazio 1992). Implementing this in a computational model requires filling in underspecified aspects of this relationship (e.g., shape and parameters of the distribution function of the effect of hunger on attention). By comparison, a validation test can be based closely on the empirical tests conducted (e.g., a correlation test between those variables can be evaluated using a simple simulation). Moreover, since the tests have already been created, they can always be run on later iterations of the cognitive agent to ensure that new relationships do not violate existing ones. If such violations occur, either some problem exists with the cognitive model implementation or, more interestingly, it might imply that two findings or theories from literature are incompatible for some reason (e.g., only valid for different sets of as-yet-unknown boundary conditions). Similar violations might occur when applying a model to a new task. For example, the task dynamics might drive the cognitive model to behave in a way that violates its specifications derived from literature. In either case, tests based on theory and empirical findings provide a “canary” in the system to detect such events quickly. This allows a modeler to correct their implementation or to look further into why certain research implies conflicting outcomes. The existence of metadata requirements for different cognitive agents also enables a limited form of model comparison. The gold standard for model comparison is probably model docking, where two models are directly applied to the same task and their behavior is compared (Burton 2003). However, except for well-formed (e.g., neat) models that allow direct inferences of equivalence, this process can be time-consuming and often requires a full experimental methodology. By comparison, comparing the metadata about requirements is much simpler. Assuming that two different sets of metadata can be represented in the same format, a number of comparisons can be made between their resulting models. First, the intersections and differences between the parameter sets of each model can be compared (e.g., which concepts do both models study?). Provided these can be expressed using a common ontology, the relative parameter sets would tell a great deal about the focus of each model. Likewise, requirement sets could be compared (e.g., how much do these models employ the same underlying assumptions about cognition?). If metadata about the scope of tasks for each model was also available, this ap- proach would allow comparing if two cognitive agents are actually two different ways to approach the same problem. Finally, as requirements are intended to be tied to papers, this approach could be used to compare the distance between literature behind different models (e.g., are they using different papers on the same theory or do they come from completely non-intersecting theoretical backgrounds?). While these comparisons cannot replace the empirical comparison between model behavior offered by model docking, they could be performed easily across any two cognitive agents that have metadata about their requirements. As such, they could detect pairs of models that could be meaningfully docked for further study, improve searching for cognitive models (e.g., “find similar models” functionality), and even cluster models by the similarity between their parameters or literature. While these capabilities would not replace the power of model docking, they would provide simple and efficient comparison functionality that could be used for semantic search, to compare different versions of the same model, or to perform high-level comparisons between the properties of different cognitive models. 6 Limitations and Future Directions For designing cognitive agents and their cognitive models, traditional paradigms such as KISS (Keep it Simple, Stupid) and KIDS (Keep it Descriptive, Stupid) fail to capture the importance and traction offered by explicitly rooting design into the underlying literature. The KIKS (Keep It Knowledgeable, Stupid) approach is proposed as a process for designing cognitive agents that outlines a structure for the modeling behaviors already used for modeling cognitive agents: selecting relevant literature, identify key parameters and relationships, implementing a computational cognitive model, and validating that model. KIKS is not proposed as a revolution to existing design practices for cognitive agents and cognitive models. To the contrary, much of the KIKS process merely outlines steps that many modelers already consider. The primary differences between this approach and existing modeling techniques currently in practice is that it explicitly calls for producing metadata about the key parameters and their relationships from literature, then using those to specify validation tests before actually making the model. These differences have benefits, as well as costs. The primary cost of KIKS is the time invested to record metadata. With that said, the value of such metadata for literature search and meta-analysis purposes alone should offset that investment. Additionally, there are already movements in related research communities (e.g., neuroimaging) toward similar metadata that would dovetail with this work. While journals have only moved in this direction slowly, fields like biology offer examples for how research can benefit from the “big data” implicit in published literature and its reported findings (Altman et al. 2008). Overall, the potential benefits for such metadata are considerable and worthwhile even without KIKS. The second cost is that KIKS asks modelers to specify their validation tests prior to coding the cognitive agent. This 68 discovered through careful scientific exploration. Issues also exist for streamlining validation tests for cognitive agents, though these problems are not specific to KIKS. In general, there has been significant debate over best practices for validating computational and agent-based models (Cooley and Solano 2011). Cognitive agents, for example, often incorporate relationships derived from statistical tests based on findings from many subjects. So then, should the a distribution of cognitive agent “subjects” be tested for similar distribution properties? Alternatively, what if the cognitive model always displays this relationship across all conditions (e.g., hunger always leads to increased attention to food)? Is this a good model because it captures the main relationship from literature, or is it a bad model because it lacks the noise empirically observed across subjects? These are fundamental philosophical questions about validating the mechanisms of cognitive models that do not, as yet, have cut-and-dry solutions. However, these should not dissuade modelers from constructing validation tests. If anything, they call for additional attention to validity-testing and the implications of what is actually being tested. Finally, the KIKS principle is entirely silent on best practices for selecting appropriate literature for designing a model and on the process of choosing how to implement the model computationally. These gaps are intentional, as the literature review and computational implementation are both deeply connected to the domain expertise of research groups. Despite these limitations and boundaries, KIKS offers basic principles and a structured process for designing cognitive agents. While there is room for improvement, KIKS focuses design specifically on the cognitive model and the knowledge behind it, rather than viewing the design process in terms of a parameter search. This represents a significant improvement over other principles commonly invoked for agent-based modeling. Additionally, KIKS accommodates the issue that cognitive agents require internal cognitive modeling as well as task-specific modeling related to their intended application (e.g., simulated experiments, virtual agents, etc.). As such, the KIKS principle can hopefully contribute to the larger discussion of how and why cognitive agents are designed. does not necessarily mean that the tests must be fully implemented, but their expected functional or statistical relationships should at least be outlined. Due to project scheduling and delivery dates, it may not always be possible to follow a test-driven design process. However, it is very much in line with the scientific method: one would not design an experiment prior to establishing criteria for the success or failure of that experiment. Why then would the design of a cognitive agent, which might be used in many simulation experiments, precede formulating tests to ensure the model works as intended? Worse, after a cognitive agent is finished, it is very easy for rigorous validation to lose out to new applications or extensions to the agent. Particularly for a predictive model, leaving validation to the end of a design cycle has dangerous implications for the quality of the model and the research it produces. Given that cognitive agent modelers should be developing validation measures at some point, KIKS suggests specifying these tests first rather than after modelers are deeply invested in a particular implementation. This paper is not intended to be the last word on KIKS. Instead, KIKS is tentatively proposed as an alternative to the KISS and KIDS principles, which do not offer guidelines to harness the knowledge behind cognitive agents. Cognitive agent design, which relies heavily on cognitive model design, has unique challenges and opportunities that KIKS attempts to accommodate. However, given the breadth of research on cognitive agents, KIKS should be significantly improved by further discussion and refinements of the suggested process. The discussion of how to tighten the links between cognitive models and published knowledge may ultimately be more important than the specific proposed design process. While a high-level process has been presented, the pragmatics of KIKS also require further specification. For example, a simple and general metadata format for representing theoretical and empirical relationships is needed to promote sharing of such metadata. Likewise, while the results of standard statistical tests (e.g., location tests) are relatively easy to represent, advanced hypothesis tests (e.g., cluster analysis) are not as standardized in their reports. Translating these into metadata might offer challenges. Theorized relationships also offer challenges for formal markup, as these might require representing functions or algorithms. While methods exist for representing these relationships (e.g., symbolic algebra languages, logical languages such as Prolog), cognitive modelers would only consider representing these relationships if they could do so intuitively and with minimal effort. With that said, many cognitive models primarily consider knowledge derived from classical statistical tests, so these issues may be less critical in practice. Secondly, the question must be asked: why focus on metadata at all? Why not focus on raw experimental data instead of research findings or theories? Particularly with moves toward big data in social science, this is an important question. However, this author would argue that the value added by cognitive researchers in conducting, framing, and interpreting their analysis is pivotal. While there may also be merit in using raw experimental observations to help design a cognitive agent, it is doubtful that raw data could replace the relationships Acknowledgements Thank you to the Office of Naval Research, whose support for intelligent systems helps make this work possible. Also, many thanks to my past and present colleagues at both the University of Pennsylvania and the University of Memphis, whose vigorous scientific discussions helped set these thoughts in motion. References Altman, R. B.; Bergman, C. M.; Blake, J.; Blaschke, C.; Cohen, A.; Gannon, F.; Grivell, L.; Hahn, U.; Hersh, W.; Hirschman, L.; Jensen, L. J.; Krallinger, M.; Mons, B.; O’Donoghue, S. I.; Peitsch, M. C.; Rebholz-Schuhmann, D.; Shatkay, H.; and Valencia, A. 2008. Text mining for biology–the way forward: opinions from leading scientists. Genome Biology 9 Suppl 2:S7. 69 Anderson, J. R. 1996. Act: A simple theory of complex cognition. American Psychologist 51(4):355–365. Axelrod, R., and Hamilton, W. D. 1981. The evolution of cooperation. Science 211(4489):1390–1396. Burton, R. M. 2003. Computational laboratories for organization science: Questions, validity and docking. Computational and Mathematical Organization Theory 9(2):91–108. Cooley, P., and Solano, E. 2011. Agent-based model (ABM) validation considerations. In The Third International Conference on Advances in System Simulation (SIMUL 2011), 126–131. Edmonds, B., and Moss, S. 2005. From KISS to KIDS – an antisimplistic modelling approach. In Davidsson, P.; Logan, B.; and Takadama, K., eds., Multi-Agent and Multi-AgentBased Simulation, 130–144. Springer. Everett III, H. 1963. Generalized Lagrange multiplier method for solving problems of optimum allocation of resources. Operations Research 11(3):399–417. Gerstein, M.; Seringhaus, M.; and Fields, S. 2007. Structured digital abstract makes text mining easy. Nature 447(7141):142. Kaelbling, L. P.; Littman, M. L.; and Moore, A. W. 1996. Reinforcement learning: A survey. Journal of Artificial Intelligence Research 4:237–285. Koedinger, K. R.; Baker, R. S. J.; Cunningham, K.; Skogsholm, A.; Leber, B.; and Stamper, J. 2011. A data repository for the EDM community: The PSLC datashop. In Romero, C.; Ventura, S.; Pechenizkiy, M.; and Baker, R. S., eds., Handbook of Educational Data Mining. Taylor and Francis. 43–55. Laird, J. E. 2008. Extending the Soar cognitive architecture. In Proceedings of the 2008 conference on Artificial General Intelligence, 224–235. Amsterdam, Netherlands: IOS Press. Minsky, M. L. 1991. Logical versus analogical or symbolic versus connectionist or neat versus scruffy. AI Magazine 12(2):34–51. Myung, J. I., and Pitt, M. A. 2010. Cognitive modeling repository. cmr.osu.edu (Retrieved: May 20, 2013). Nakayama, T.; Hirai, N.; Yamazaki, S.; and Naito, M. 2005. Adoption of structured abstracts by general medical journals and format for a structured abstract. Journal of the Medical Library Association 93(2):237–42. Newell, A. 1994. Unified theories of cognition. Cambridge, MA: Harvard University Press. Poldrack, R. A.; Kittur, A.; Kalar, D.; Miller, E.; Seppa, C.; Gil, Y.; Parker, D. S.; Sabb, F. W.; and Bilder, R. M. 2011. The cognitive atlas: toward a knowledge foundation for cognitive neuroscience. Frontiers in Neuroinformatics 5(17). Rand, D. G. 2012. The promise of Mechanical Turk: how online labor markets can help theorists run behavioral experiments. Journal of Theoretical Biology 299:172–179. Roskos-Ewoldsen, D. R., and Fazio, R. H. 1992. On the orienting value of attitudes: Attitude accessibility as a deter- minant of an object’s attraction of visual attention. Journal of Personality and Social Psychology 63(2):198–211. Silverman, B. G.; Might, R.; Dubois, R.; Shin, H.; Johns, M.; and Weaver, R. 2001. Toward a human behavior models anthology for synthetic agent development. In 10th Conference on Computer Generated Forces and Behavioral Representation, 277–285. Silverman, B. G.; Johns, M.; Cornwell, J. B.; and O’Brien, K. 2006. Human behavior models for agents in simulators and games: Part I: Enabling science with PMFserv. Presence: Teleoperators and Virtual Environments 15(2):139– 162. Silverman, B. G. 2010. Systems social science: A design inquiry approach for stabilization and reconstruction of social systems. Intelligent Decision Technologies 4(1):51–74. Sun, R. 2006. Cognition and Multi-Agent Interaction: From Cognitive Modeling to Social Simulation. Cambridge University Press. Sun, R. 2007. Cognitive social simulation incorporating cognitive architectures. IEEE Intelligent Systems 22(5):33– 39. Turner, J. A., and Laird, A. R. 2011. The cognitive paradigm ontology: Design and application. Neuroinformatics 10(1):57–66. 70