| |
ISSN 1320-0682 | |||
| Volume 03 | April 1996 | |||
Pierre Marcenac
The main issue tackled in this paper is auto-adaptation - that is generation of behaviours from the study of local interactions between agents and the environment. Some key issues associated with the understanding and representation of such emerging behaviours in multi-agent systems are introduced.
Keywords:
Emergence, Local Interactions, Complex Systems Modelling,
Agent-Oriented Knowledge Representation, Simulation Applications.
The general framework of the project is the study of the agent paradigm and its intrinsic mechanisms in order to model complex systems. A complex system can be defined as a system in which any algorithm could describe the behaviour, and where mathematical models do not provide efficient solutions to understand and then to predict underlying phenomena. One of the privileged applications of such a framework is natural phenomena modelling. The issue is not to model some existing systems which are naturally distributed, but rather to model complex systems with artificial systems. This point of view can be, at a first glance, associated to C. Langton's works on artificial life [5].
In our project, a simulation application to help in predicting volcanic eruptions has been investigated. To try to understand complex behaviours, a computational model with communicating agents is then considered, in which emergent phenomena arise through interactions between local entities and their environment. Such a multi-agent system is described in a layered architecture, called GEAMAS (acronym for GEneric Architecture for MultiAgent Simulations). GEAMAS is seen as a multi-agent software platform, intended to develop simulation applications. Envisaging a layered architecture allows a better understanding of how the emergence of a global behaviour occurs, and why the multi-agent approach works in this context. This architecture is based on three abstraction levels. Each successive level represents a higher level of abstraction, and describes a complex degree of knowledge.
The article introduces some key issues associated with the understanding and representation of the emergence of behaviours in the architecture. When modelling complex systems, one of the characteristics is that behaviours can not be linear during time and dynamically evolve during the simulation. Evolution capabilities should then be given to agents when designing the system. Such evolution capabilities can be classified as:
The paper focusses on the first point - that is, the adaptation of behaviours according to knowledge and environment evolution. This adaptation is seen in GEAMAS as the result of interactions between agents and their environment. To give the architecture the means to display such a functionality, a specific message type, the "Recomposition", has been introduced. The Recomposition mechanism allows information transfer to higher levels in the architecture when instability of reactive agents is detected.
The paper is divided into two main parts. The next section presents the architecture composed of the society, cognitive and reactive agents. The following describes how complex behaviours occurs from the study of local interactions between reactive agents in such a context. Some elements of the implementation are briefly covered. Finally, the last section concludes the article by pointing out current investigations.
In complex systems modelling, two kinds of features are to be considered. On the one hand, non-predicted and complex real worlds need a distribution of the complexity in atomic pieces to better understand, control, and isolate underlying phenomena. On the other hand, agents could not model their surroundings in details, as there are too many unknown events to consider. It is then not surprising that neither purely reactive nor purely cognitive architectures can be flexible enough for complex systems modelling. This fact is widely accepted in many areas, related in Fergusson (1994) [2] and Wooldridge and Jennings (1994) [10] for instance. The approach proposed in GEAMAS is to marry cognitive and reactive architectures. Such an architecture is called hybrid, according to the common definition found in Wooldridge and Jennings (1994) [10].
The architecture of GEAMAS describes an abstracted model of the real world.
This foundation is inherent to the problem to be solved, namely modelling and
distributing complexity. To find the adequate separation, GEAMAS is based on
the distribution of roles: macro-agents are related elements in which emergent
behaviours will be observed, and describe agents in an organised society;
interactions, micro-behaviours and evolution facilities of the agent are
embedded into cognitive and reactive agents.
In GEAMAS, the society describes the organisation of the agent's structure composing the system. Such a structure is represented in a network of acquaintances, forming a competence network, where agents correspond to nodes and interaction possibilities (acquaintances) to edges. A network of acquaintances is assumed to be one of the easiest ways to implement connections between agents - that is, the number of acquaintances for each agent. When the simulation begins to run, each agent composing the society establishes his acquaintances (represented by his neighbours in the network), to be able to communicate with other agents asynchronously.
The number of acquaintances constitutes a very interesting feature, because it addresses the way a real world could be designed as a three-dimensional network [7]. This parameter is called the degree of communication or the connectivity of the agents in the system. As connections are determining the ability of an agent to communicate, a part of the communication protocol is then defined by setting this parameter. This number of connections describes the influence an agent could apply on another, and defines a spatial and geometric disposition of agents in the universe. It is then possible to abstract this projection, showing then a view of the real world in three dimensions, by considering 3-D connections in a network. A plan then becomes an abstracted view of the real world, and if two agents are not immediate neighbours, they can even be linked together. Note, however, that if agents are moving from one place to another, the network is re-organised, involving the agents' connectivity being updated at the same time.
Each cognitive agent is described by the triad - interaction, behaviour, and evolution:
Reactive agents constitute the last level of the architecture. Interactions between reactive agents are provided by signals, as stimuli-reactions. Signals do not bear semantics, and their meaning depends on the interpretative ability of the receiver. Some parameters define the agent's state, and are called state parameters. An agent is assumed to be stable when the values of his state parameters do not hit thresholds (that is, critical values). Thresholds and state parameters are given by the designer before the simulation begins.
In a universe describing natural phenomena for instance, each component is subject to nonlinear modifications during time. Modifications affect both agents' data and behaviour. Two kinds of dynamic evolution could be considered: modification of the structure by re-organisation of the acquaintances network or modification of behaviour.
The first one constitutes a fundamental feature in complex systems modelling, and is often called self-organisation. Indeed, this property allows a system to be simulated without any initial constraints. The network is then organised as events arrive, until the system hits a critical state. In most natural systems, the critical state is the most favourable state to observe global phenomena. This part of the work is beyond the scope of this paper.
The modification of behaviour is a fundamental notion for an agent. This
characteristic authorises an agent to reproduce adequate behaviours - that is,
those behaviours most adapted to his environment and the current situation.
Multi-agent systems typically are of considerable complexity with respect to
both their structure and their functionality. It is extremely difficult to
correctly determine when and why a behaviour of an agent or an activity of a
multi-agent system occur at the time of its design. This kind of problem can be
reduced by giving the agent the ability to adapt himself to new situations.
This feature will be called self-adaptation so far. Self-adaptation is
then defined as self-modification that enables a system to survive in a changed
environment [9].
Understanding the reasons for which behaviours emerge is often surprising and
seen as a kind of "magic phenomenon" in multi-agent systems
[1].
Nevertheless, emerging behaviours are the result of interactions and situations
of co-operation between underlying agents in the architecture. Co-operation
between agents encourages emerging behaviours, because collective agents' works
could make an action possible.
Emergence in GEAMAS has been identified through a specific mechanism, the "Recomposition", which has been isolated in the architecture. Recomposition transfers information from reactive agents to cognitive agents. A Recomposition message is used by reactive agents to express their instability. The agent's instability is given by critical values (thresholds) of state parameters. The values of state parameters evolve during the simulation, according to multiple local interactions between reactive agents. When a threshold is hit, the agent is assumed to be unstable, and information is then transferred to the cognitive agent with a Recomposition message.
Hence, a Recomposition message is provided by reactive agents to alert cognitive agents and respectively the global society of agents, that something unusual happens. Shifting from the micro-level up to the upper-level, cognitive agent or society collects data on micro-behaviours and combines them to determine macro-behaviour and adapts himself to the situation. A higher behaviour is then emerging from this adaptation. In such a sense, interactions between reactive agents set emerging behaviours.
The architecture of GEAMAS is actually implemented in an agent language, ReActalk [3], that is an agent environment to develop and experiment with agents. The notion of self-adaptation has been realised thanks to reflection. Reflection allows an agent to draw a feedback of his behaviour and to dynamically update and modify it. For our proposal, the reflection mechanism helps in implementing the interpretation of the instability of reactive agents, when a Recomposition message is received by an agent. The agent receiving the message must then respond by adapting his behaviour. Such an agent is called a reflective agent.
Reflection then provides a natural framework for expressing dynamic evolution of agents behaviours. When an agent is created, a meta-agent is given to it. This meta-agent provides an easy access to the self-adaptation facilities of an agent. The whole structure of the system is described in Giroux et al. (1996) [4], and the ReActalk environment in Giroux (1996) [3].
The emergence mechanism presented in this paper has been experimented in the frame of the simulation of volcano behaviours. To better understand the mechanics behind volcanoes, a simulation platform based on GEAMAS principles was designed and implemented [7]. Micro-behaviours encapsulating the complexity were defined and assigned to reactive agents. The focus was put on the circulation of magma through rocks. The whole application was used over twelve months by our team, producing more than one hundred simulations. The interpretation of such primary results was very encouraging, as some critical parameters playing a role in eruptions were raised. Simulations have investigated processes that can neither be measured on volcanoes nor be modelled using classical linear mechanics [8]. Another application of the architecture, described in Leman et al.(1996) [6], has been performed for Intelligent Tutoring Systems. We are now convinced that this platform is appropriate for simulation of other natural systems exhibiting similar behaviours, such as natural phenomena, risk prediction, social behaviour analysis, etc.
Actual investigations concern the integration of machine-learning to tackle the emergence of non-predictable behaviours for prediction applications.
The author is particularly grateful to Sylvain Giroux for his involvement in the project during his 1993-1994 post-doctoral studies, and those people who have brought their contribution to this work, namely S. Calderoni, J.R. Grasso and D. Grosser.
Emergence of Behaviours in Natural Phenomena Agent-Simulation
This document was generated using the LaTeX2HTML translator Version 96.1 (Feb 5, 1996) Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
The command line arguments were:
latex2html cs964.
The translation was initiated by Pam Milliken on Tue Jan 21 11:44:36 EST 1997