Self-Oraganization_M.. - LES PUC-Rio

advertisement
Self-organization in Multi-Agent Systems
Manoel Teixeira de Abreu
mnetto@les.inf.puc-rio.br
Roadmap
• Motivation
– Multi-agent System
• Self-Organization
– Definitions
– Properties
– Characteristics
– Measures
• Emergence
– Definitions
– What Does Emerge?
– Properties
• Design Patterns for Decentralised Coordination
Motivation
• Operate in dynamic, heterogeneous environments and face
the challenge of handling frequently changing requirements.
Such ubiquitous computing, biological systems and
autonomic computing.
• Must be flexible, robust and capable of adapting to the
circumstances.
• The ability of a system to have a complex collective
response arising from interactions among relatively simple
individual components.
Multi-agent System
• Autonomous software entites are agents
• MAS is a group of cooperating agents
• Macroscopic behaviour using only locally agents interations
Roadmap
• Motivation
– Multi-agent System
• Self-Organization
– Definitions
– Properties
– Characteristics
– Measures
• Emergence
– Definitions
– What Does Emerge?
– Properties
• Design Patterns for Decentralised Coordination
Self-Organization - Definitions
• Defined as the mechanism or the process enabling a system
to change its organization without explicit external
command during its execution time.
• Strong self-organizing systems are those systems where
there is no explicit central control either internal or external.
Purely decentralized (access to global information is limited
or impossible)
• Weak self-organizing system are those systems where,
from an internal point of view, there is re-organization
maybe under an internal (central) control or planning.
• Interactions occur locally (among neighbours) and based on
local information.
6
Self-Organization – Properties (I/II)
• Absence of explicit external control. This is a mandatory
property that states that the system is autonomous; it
imposes and changes its organization based solely on
internal decisions and without following any explicit external
(re-) organization command. This property refers to the
self- part of the self-organization definition.
• Decentralized control. A self-organizing system can work
under decentralized control. In this case, there is no internal
central authority or centralized information flow. As a result,
access to global information is limited by the locality of
interactions, which is governed by simple rules. This
property is generally not mandatory, since we can observe
internal central control in many natural self-organizing
systems, such as the termite systems.
Self-Organization – Properties (II/II)
• Dynamic operation. This mandatory property is linked to
the system evolution in time. Since the organization evolves
independently of any external control, this property implies
continuity in the self-organization process.
Self-Organization – Characteristics (I/III)
• Endogenous global order. The system reaches some (stable)
global state that is produced from within the system.
• Emergence. This refers to having emergent phenomena arising
from local interactions. Such phenomena can be observed at global
level but they cannot be accurately deduced from examining
individual behaviours. Arise from local interactions occurring
among the individual components.
• Local rules. The overall complex system behaviour can be based
on simple individual behavioural rules. Local information describes
the mechanism for producing the global behavioural pattern and
not the pattern itself.
Self-Organization – Characteristics (II/III)
• Dissipation. In the absence of external perturbation, the
system is expected to stabilize in some states in which
emergent properties can be observed. This implies a kind of
dissipation of some ‘energy’, otherwise the system would be
continuously changing.
• Instability. Systems showing instability are mainly
characterized by nonlinear dynamics which have the effect
of small fluctuations in environmental conditions that result
in significant variations in overall system behaviour.
• Multiple equilibria. Multiple equilibria are observed when
many possible attractors for stable states are present in the
system.
Self-Organization – Characteristics (III/III)
• Redundancy. A system presents redundancy when it
demonstrates insensitivity to damage due to replication of
components.
• Self-maintenance. A system is self-maintained if it has the
capacity to reproduce or repair itself, essentially by reproducing or
repairing its components.
• Adaptivity. The re-organization capability of self-organizing
systems implies adaptation to external (environmental) variations.
• Complexity. This characteristic usually arises from the
irreducibility of the global properties to a combination of local
behaviours.
• Hierarchies. Hierarchies are present in a system when multiple
nested self-organized levels can be observed.
Self-Organization - Measures
Typical self-organization measures includes:
• Capacity to reach an organization able to fulfil the goal of
the system as a whole once the system is started
(success/failure/time required, convergence);
• Capacity to reach a re-organization after a pertubing event
(success/failure/time required);
• Degree of decentralized control (central/totally
decentralized/hybrid);
• Capacity to withstand pertubations (stability/adaptability)
Roadmap
• Motivation
– Multi-agent System
• Self-Organization
– Definitions
– Properties
– Characteristics
– Measures
• Emergence
– Definitions
– What Does Emerge?
– Properties
• Design Patterns for Decentralised Coordination
• Some examples
Emergence - Definitions
• “The whole is more than the sum of the parts”. Witch
captures the essence of the emergent phenomena.
• Tom de Wolf et al., 2005 [1] defines Emergence as “A
system exhibits emergence when there are coherent
emergents at the macroscopic level that arise from the
interactions between the parts at the microscopic level.
Such emergents are novel w.r.t. the individual parts of the
system.”
• Di Marzo Serugendo el at., 2006: “A structure (pattern,
property or function), not explicitly represented at the level
of the individual components (lower level), and which
appears at the level of the system (higher level).”
14
Pheromone Path of Ants
• What does emerge?
• It is considered emergent if and only if there exists an
observer capable of observing and characterizing it as such.
15
Emergence – Properties (I/II)
• Novelty. Refers to the fact that although the resulting
phenomenon is derived from a particular organization and
interaction of the micro-level parts, it is radically different
from their individual properties and it cannot be derived or
predicted from them.
• Irreducibility. Structures and/or functions appear at a
macroscopic level and the observation of the component
properties cannot enable to predict them (Atlan, 1993). For
example, in complex systems, the complexity often results
in features, emergent properties, which are properties of the
system that the individual parts do not have.
Emergence – Properties (II/II)
• Interdependency between levels. In these systems
there are at least two levels, a micro-level corresponding to
the substrate where the emergent phenomena comes from
and a macro-level enabling the observation of the emergent
phenomena. The micro-level causes the emergent
phenomena and the macro-level constrains the behaviour of
the entities at the micro-level. Therefore, there is a strong
dependency between the dynamics observed at both macroand micro-levels.
• Nonlinearity. An emergent phenomenon originates from
nonlinear activities at the micro-level. Typical examples of
nonlinear activities are loops of positive and negative
feedback.
Why Decentralized?
• Computation and decisions to be distributed among the
different components, thus preventing the need for a central
powerful computer;
• The system is more robust since it does not rely on a single
node that may fail and crash the whole system;
• Network and CPU resources are better used in the sense
that communication does not occur among a dedicated
central node and a large number of components, but locally
among the whole set of components;
• In dynamic systems, where components join and leave the
system permanently, decentralized control allows a flexible
schema for communication, e.g. with a neighbor instead of
with the central entity.
Roadmap
• Motivation
– Multi-agent System
• Self-Organization
– Definitions
– Properties
– Characteristics
– Measures
• Emergence
– Definitions
– What Does Emerge?
– Properties
• Design Patterns for Decentralised Coordination
Design Patterns for Decentralized
Coordination
• Choice of one or more decentralised coordination
mechanisms to achieve the desired macroscopic behaviour.
• Using such a set of structured patterns, self-organising
emergent system can be designed more systematically.
Gradient Fields
• Context/Applicability
– Need for decentralized coordination of multiple autonomous
entities in a environment which is robust and flexible in face of
changes.
• Problem
– Routing of messages, agents.
• Solution
– Come from physics: movement of masses and particles
according to gravitational/electromagnetic fields
Gradient Fields
• How does the mechanism work?
22
Gradient Fields
• Solution (cont.)
– Parameter Tuning:
• Propagation factor, propagation rate -> How quick are updates
needed when changes occur
– Infrastructure:
• What is need to support the mechanism? An environment
supporting field storage and propagation
– Characteristics:
• Handles dynamic situations – Robust
• Simple agents – Complex Environment. Environment transparently
and completely handles coordination issues.
• Know Uses
– Spatial shape formation
– Urban Traffic Management
23
Problem(s)
Solution
Pattern
Spatial Source to
Destination Routing,
Task Recruitment,
Relation Identification,
Integration of various
information Sources
Agents explicitly search for goals, tasks, or related items and drop
pheromones to form paths for other agents to follow to the goal or task.
Reinforcement of an existing path by other agents can be seen as a
reinforcement of the relation between source and destination.
Evaporation, Aggregation, and Propagation keep the pheromones up-todate and support integration of various information sources.
Digital
Pheromone
Paths
Spatial Movement,
Pattern Formation,
Structure Formation,
Routing, Integration of
Contextual Information
Spatial, contextual, and coordination-related information is automatically
spread/propagated by the environment as multiple computational fields.
Agents simply follow the “waveform” of these fields to achieve the
coordination task, no explicit exploration is needed. The spatial
information comes to the agents instead of agents explicitly searching.
Gradient Fields
Resource Allocation in
general (resource=task,
power, bandwidth,
space, time, etc.) ,
Integration of resource
Usage/Need Information
A virtual market where resource users sell and buy resource usage with
virtual currency. The price evolves according to the market dynamics and
indicates a high (high price) or low (low price) demand. This information is
used by agents to decide on using the resource or not. Economic market
theory states that the prices converge to a stable equilibrium.
Market-based
Coordination
Team-formation, Trust
and reputation
Agents put and modify tags on other agents and a team is formed by only
collaborating with agents with the same tag or some other condition. If
tags indicate how well agents behaved in collaborations with others then
trust and reputation information can be available.
Tags
Synchronisation,
Resource Allocation
A capability or a resource is represented by a token. Synchronisation
happens because only the holder can use the resource or executes a
certain capability. Tokens can be passed among agents to allocate the
resource or capability to others.
Tokens
…
…
…
Simple Patterns
• Replication
– Allows for recovering minor mutations;
– Lowers the time required to access a resource.
• Collective Sort
– Grouping together related information helps to manage batch
processing;
– The goal of Collective Sort is to group together similar information in
the same node, while separating different kinds of information.
• Evaporation
– Mechanism to reduce information amount, based on a time relevance
criterion.
• Aggregation
– Mechanism of reinforcement (eg. pheromone).
• Diffusion
– When pheromone is deposited into the environment it spontaneously
tends to diffuse to neighboring locations.
References
•
Gardelli, Luca and Viroli, Mirko and Omicini, Andrea. "Design Patterns for Self-Organizing
Multiagent Systems”, 2nd International Workshop on Engineering Emergence in Decentralised
Autonomic Systems (EEDAS 2007), CMS Press, University of Greenwich, London, UK. 61-70.
•
G. Di Marzo Serugendo, M.-P. Gleizes, A. Karageorgos. "Self-organisation and emergence in
MAS: an overview", Informatica 30(1): 45-54, Slovene Society Informatika, Ljubljana, Slovenia,
2006.
•
G. Di Marzo Serugendo, M.-P. Gleizes, A. Karageorgos. "Self-Organisation in MAS", Knowledge
Engineering Review 20(2):165-189, Cambridge University Press, 2005.
•
G. Di Marzo Serugendo. "Autonomous Systems with Emergent Behaviour". Chapter in
“Handbook of Research on Nature Inspired Computing for Economy and Management”. Jean-Philippe
Rennard (Ed), Idea Group, Inc., Hershey-PA, USA, pp. 429-443, September 2006.
•
T. De Wolf, Analysing and engineering self-organising emergent applications, Ph.D. Thesis,
Department of Computer Science, K.U.Leuven, Leuven, Belgium, May, 2007.
•
T. De Wolf, and T. Holvoet, Emergence Versus Self-Organisation: Different Concepts but
Promising When Combined, Engineering Self Organising Systems: Methodologies and
Applications (Brueckner, S. and Di Marzo Serugendo, G. and Karageorgos, A. and Nagpal, R., eds.),
Lecture Notes in Computer Science, 2005, Volume 3464, May 2005.
•
T. De Wolf, and T. Holvoet, Design Patterns for Decentralised Coordination in Self-organising
Emergent Systems, Editors: Sven Brueckner, Salima Hassas, Màrk Jelasity and Daniel Yamins,
Engineering Self-Organising Systems: Fourth International Workshop, ESOA 2006, Future
University-Hakodate, Japan, 2006, Revised Selected Papers, Lecture Notes in Computer Science,
Volume 4335, 2007, pp. 28–49, Springer Verlag
26
References
•
Jan Sudeikat and Wolfgang Renz: "Toward Requirements Engineering for Self-Organizing
Multi-Agent Systems" in: Proceedings of the First International Conference on Self-Adaptive and
Self-Organizing Systems, 2007.
•
Holland, J. H. (1998). Emergence – from Chaos to Order. Oxford University Press.
The End.
Questions?
Download