An environment-based principle to design reactive multi-agent systems for problem solving. Olivier Simonin1 and Franck Gechter1. Universite de Technologie ...
An environment-based principle to design reactive multi-agent systems for problem solving Olivier Simonin1 and Franck Gechter1 Universite de Technologie de Belfort-Montbeliard, 90010 Belfort cedex, France, {olivier.simonin}{franck.gechter}@utbm.fr, WWW home page: http://set.utbm.fr/membres/simonin/
Abstract. Even if the multi-agent paradigm has been evolving for fifteen years, the development of concrete methods for problem solving remains a major challenge. This paper focuss on reactive multi-agent systems because they provide interesting properties such as adaptability and robustness. In particular, the role of the environment, which is the real place where the system computes and communicates, is studied. From this analysis a principle to design or engineer reactive systems is introduced. Our approach is based on the representation of the problem’s constraints considered as perturbations to stabilize. Finally, the relevancy of our proposal is justified through the development of two solving models applied to real and complex problems.
1
Introduction
Even if the multi-agent paradigm has been evolving for fifteen years, the development of concrete methods for problem solving remains a major challenge. This paper tackles this problem by providing a first step towards a methodology aimed at designing reactive multi-agent solutions. It focuses on reactive systems because they present interesting features as self-organization/emergent phenomena, robustness, adaptability, simplicity and redundancy of the agents (then low cost agent design). It has been shown that this approach is efficient to tackle complex problems as life-systems simulation/study [7], cooperation of situated agents/robots [9], problem/game solving [3],... However, it is difficult to extract a generic method to build a reactive-based solution facing a (distributed) problem. This difficulty is due to the complexity of such systems where agents and interactions are numerous and where global dynamics are complex to control. As it has been emphasized in [12] and [10], the environment plays an important role in reactive multi-agent systems (MAS). It is the main place where the system computes and communicates. In the problem solving framework, it is clear that a reactive agent can neither handle a representation of the problem nor compute its solution. Such systems are based on many agent-agent and agent-environment interactions. As agents’ interactions are reactions to perceptions, it appears clearly that they are usefull to enable solving dynamics, but they do not provide a mean to express the problem. So, the problem representation can only be defined in the environment model. This paper presents how
to establish the link between the problem’s representation, expressed as environmental constraints, and the agents’ behaviors which are regulation items of the environmental perturbations. This method contrasts with classic approaches that consists to define agents and interactions following the structures we expect to see to emerge (as proposed in [9] [10]). Section 2 defines an environment-based solving principle for reactive MAS. Then, section 3 illustrates our approach on two applications. Finally, section 4 concludes on this proposition and presents future work.
2
Proposition of a methodology to build an environment-based solving system
As opposed to the socio or bio inspired approaches, we propose a more pragmatic engineering method to define reactive-based problem solving systems. Our approach is closely tied to the standard regulation loop defined in automatic control. The goal of the problem solving is to build a solution, stable in time and space, considering the formulation of a problem that has its own topology1 and dynamics2 . Thus, the MAS can be considered as a regulation (or filtering) process. As a consequence, solving a problem leads to define the parameters of the regulation loop in order to obtain a stable output (solution level) considering the variations of the input (problem level). (see figure 1)
Interactions Perturbations Problem to solve
Perceptions ENV.
=
Pb. Model
Agents
Result / Organisation = State of the System
agent model real and/or virtual world
Fig. 1. Environment based solving principle
The environment is defined as the input layer of the regulation loop. It translates the variations and the topology of the problem and presents them to the agents. The organization is the output layer of the system. It represents the state of the system on spatial and temporal level. The regulation mechanism is defined by the agents’ actions and their interactions. These interactions have been divided into two categories. The first one characterizes the agent-agent 1 2
i.e. how the problem is structured in space i.e. how the problem evolves
interactions which composes the direct branch of the regulation loop that is considered, in automatic control, as an amplification (positive feedback [10]). The negative feedback of the regulation loop is realized by the agents-environment interactions. The environment is modified by the problem and the regulation dynamics. As opposed to the problem’s one, the regulation dynamics can be controlled through the MAS parameters. The proposed principle can be described into four main steps: 1. Define an environment model which represents the problem that has to be solved on both topological and dynamical levels. Note that the problem’s model can be totally virtual or includes parts of the real world. 2. Define the agents perceptions. Agents must be able to perceive the environment’s state (perception of the perturbations that trigger their reaction). 3. Define agents interaction mechanisms in order to reduce the perturbations. These mechanisms can be defined following 3 levels (i) Provide individual and local reactions to the perceived constraints (ii) If individual actions are not efficient, provide local actions/interactions that enable cooperative processes (iii) Provide actions to regulate the previous processes. 4. Observing the result as an emergent structure, in terms of agents or environment as defined in [11]. This structure is the consequence of dynamics of the solving principle. This result can only be observed by an external viewer, which can be the designer or a particular agent of the system. In this principle, the environment is the place that links the problem and its constraints to the solving process. Moreover, it participates to the solving process thanks to its dynamical and integrative properties. Indeed, it integrates the problem and the solving process dynamics.
3
Application
This section presents two applications following the four steps of the proposed principle. 3.1
The satisfaction-altruism model
This model allows to combine self-organization and individual intelligent behaviors in situated MAS. To introduce local intelligent interactions into the collective approach, this model proposes an extension of the artificial potential fields (APF) approach. The APF technique relies on agents perceptions which allow to combine obstacles avoidance and attraction to goal [1]. 1. In order to express agent intentions and constraints, the satisfaction-altruism model [13] introduces new dynamical fields in the environment. They are defined as attractive or repulsive signals broadcasted by agents to their neighbors (see fig. 2.a). 2. Agents perceive theses new constraints through a simple signals emissionreception module and they can combine them with other perceptions.
Agent
a
perception
b
of neighbors
c
Signals
action selection & emissions
neighbor agents
Environment a.
neighbors evaluation
perception
Actions
b.
Fig. 2. (a) Scheme of agent-agents and agent-environment interactions in the Satisfaction-Altruism model (b) Conflictual situation between individual robots
3. One interesting result of this model is its ability to solve access conflicts between several robots/agents blocked in narrow passages (see such a situation in fig. 2b). Following our principle, the local perception of obstacles is used as stimulus to avoid them (agent-env. interaction, level 1). If agents are blocked, due to the presence of others, this local avoidance behavior can be inefficient. Then a mechanism of level 2, an agent-agent interaction, based on the emission of repulsive signals is added. Agents measure their local environment constraints to emit a level of dissatisfaction (agents and walls have not the same weight in this computation, see [13]). The altruistic reaction, which is a cooperative behavior, forces the less dissatisfied agents, i.e. the less constrained, to move in order to unlock the situation. This reaction is propagated to all agents involved in the blocking. As a consequence, an emergent global freeing movement appears (see details in [8]). However, such an aggressing principle can lead to oscillations and cycles. To avoid them, a blocked agent emits its initial dissatisfaction while it is not free. This notion of persistence, mechanism of level 3, allows to avoid oscillations between a constrained state and a partially solved state. 4. The observed solution is then a dynamical stability found between the problem dynamics and the agents interaction one. This model has been experimented on different simulated problems such as collaborative foraging, navigation in constrained environments, box-pushing [2] and validated with real robots in conflict problem solving (see [8]). 3.2
A Physics based reactive model
The localization problem can be defined as the fact of finding the position of an object, mobile or not, in a well known referential. As for the target tracking, it can be considered to be a series of localizations temporally and spatially coherent. The algorithms used stem generally from the signal processing or stochastic methods. Considering the principle previously enounced, a Physics based reactive model has been designed to tackle with this problem (cf. figure 3).
Fig. 3. Architecture of the Physics based reactive model for the localization and the tracking (left) and representation of the solving process as a filter (right).
1. The agents’ environment is an occupancy grid that represents of the real world (1 state ⇐⇒ 1 possible position for a target). The real targets appear, move and disappear . Thus, the environment’s dynamics has to represent these three phenomena. These are translated into two main trends: an accumulation of percepts that corresponds the appearing and an evaporation of them that deals with the disappearing. The movements are materialized by the combination of the two trends. The accumulation and evaporation mechanisms can be viewed as perturbations linked to the dynamics of the problem. 2. The agents are defined to be mass particles in a force field generated by the physics of the environment (attraction for the percepts / repulsion between agents). 3. A fluid friction has been designed to make the environment dissipative. In order to regulate the population of agents on a percept, a consumption behavior has been introduced. The perception of the perturbations of the environment is made indirectly by the attraction/repulsion behaviors. The regulation is performed through the consumption and the evaporation. (Details on the design of the interaction mechanisms can be found in [6]) 4. The collective emerging organization is both a gathering of the agents on the percepts, that leads to a group construction, and a homogenous repartition of them in the percept-less areas. Each group can thus be considered as a localized target. The output of the system is stable when an equilibrium between the refreshing and the resolution dynamics is established. The global structure of the developed device is shown in figure 3. On the application point of view, this device has been successfully applied on both simulation and with real targets. It shows relevant results compared with classical localization and tracking algorithms (see [6] or [5] for detailed results).
4
Conclusion
This paper presents an environment based principle for building reactive multiagent systems aimed at dealing with the problem solving issue. In the described approach, the environment has a key role. First, it modelizes the problem to solve and its contraints. Second, it establishes the link between the problem’s world on one side and the reactive solving process on the other. Our approach constrasts with usual emergentist or artificial life works that consist to define agents and interactions following the expected emergent organization. Our proposition can be seen as a bottom-up methodology based on the problem’s representation, where constraints are translated into perturbations in the environment. These have to be regulated through agents’ behaviors. This model is actually applied to features extraction in image processing and to robots soccer control. On the theoretical level, we plan to refine the environment-based design principle to propose a methodology.
References 1. Arkin R.C. Behavior Based Robotics The MIT Press, 1998 2. Chapelle J, Simonin O, Ferber J. How Situated Agents can Learn to Cooperate by Monitoring their Neighbors’ Satisfaction Proc. 15th European Conference on Artificial Intelligence, pp 68-72, 2002 3. Drogoul A., Ferber J., Jacopin E. Pengi: Applying Eco-Problem-Solving for Behavior Modelling in an Abstract Eco-System in Modelling and Simulation: Proceedings of ESM’91, Simulation Councils, Copenhague, pp. 337-342, 1991 4. Ferber J., Multi-Agent System: An Introduction to Distributed Artificial Intelligence. Harlow: Addison Wesley Longman, 1999. 5. Gechter F. and Chevrier V. and Charpillet F., A Reactive Multi-Agent System for Localization and Tracking in Mobile. In 16th IEEE International Conference on Tools with Artificial Intelligence - ICTAI’2004, 2004. 6. Gechter F. and Chevrier V. and Charpillet F., Localizing and Tracking Targets with a Reactive Multi-Agent System. In Second European Workshop on MultiAgent Systems - EUMAS’04. (Barcelona, Spain), 2004 7. Kennedy J. and Eberhart R.C., Swarm Intelligence, Morgan Kaufmann Publisher 2001 ISBN 1-55860-595-9 8. Lucidarme P, Simonin O, Liegeois A. Implementation and Evaluation of a Satisfaction/Altruism Based Architecture for Multi-Robot Systems Proc. IEEE Int. Conf. on Robotics and Automation, pp. 1007-1012, 2002 9. Mataric M. J., Designing and Understanding Adaptative Group Behavior Adaptive Behavior 4:1, pp 51-80, 1995. 10. M¨ uller J-P. and Parunak H.V.D., Multi-Agent systems and manufacturing INCOM’98, Nancy/Metz, 1998 11. M.R.Jean, Emergence et SMA, 5eme Journ´ees Francophones sur l’Intelligence Artificielle Distribu´ee et les Syst`emes Multi-Agents, AFCET, AFIA, La Colle-surLoup, Quinqueton, Thomas, Trousse (eds), pp 323-342. 1997 12. Parunak H.V.D., Go to the Ant: Engineering Principles from Natural Agent Systems. . Annals of Operations Research, 1997. 13. Simonin O, Ferber J. Modeling Self Satisfaction and Altruism to handle Action Selection and Reactive Cooperation in proceedings SAB 2000, v. 2, p 314-323, 2000
