Key words: Ecosystems, Multi-Agent Systems, Homogeneous vs. Heterogeneous Agents, Adaptation, Co-operation, Soccer and Learning.
Introduction
This work attempts to study the impact of Multi-Agent diversity in a
co-operative domain as soccer. Her heterogeneity is achieved by means of
having agents with different physical abilities. We plan to study some
properties in homogeneous and heterogeneous teams of
physical agents, These properties are team performance, adaptability,....
and how these properties (speed, performance,..) change depending on the
degree of diversity. We will approach the problem of defining these heterogeneous
systems by means of an ecosystem formulation, this approach is a great
help for creating hierarchical relationships among the agents, as in a
natural ecosystem, in which some species are better adapted to environment
than others.
Approach
In order to model iterations among agents, we have built our model on [1]. In this work the iterations among a large number of agents have been modelled by means of an ecosystem. For our purpose we have used this model and we have adapted it to our approach. As [1] our agents compete for bounded-resources, in our case these resources are roles in a soccer team. Each of these roles contains a limited number of actions. As agents use these resources, they get rewards from the actions they take, depending on the level of accomplishment and performance. After a given time tw, all rewards for agent s and resource r are added and an expected value is computed Grs, and frs values (also called populations) are updated according to the ecosystems equations:
These agents can be seen as economic agents that atempt to maximise an utility function. This utility function in this work are the rewards the agent gets as i interacts with the environment. This agent expects to get more rewards using past experiences and using the available system informtion, agents comunicate reliable information to each other every tw, and they are aware of the scored/received goals. This interaction is modeled by means of this ecosystem. As we said, these agents are in a environment with bounded resources and they compete for these resources with other agents. As a result of this process agents have conflicts among them, as they try to use the same action. When two or more agents attempt to make use of the same action they start a consensus process in order to decide which one will use the action. This process is conducted by means of a consensus algorithm. This algorithm uses population values and a certainty value, as team roles are coded by fuzzy sets and variables, in order to compute a value associated for each of the conflicting agents. This population value depends on every agent individual preferences and how this agents are adapted to environment. Agents better adapted to the environment, higher rewards, will tend to win more conflicts than other agents less adapted. These interaction with the environment, initially showing a chaotic behaviour, will evolve to a self-organised system where each agent will use one or more resources depending on its own physical abilities and the environment, other agent abilities and the opponent. Againt hard opponents they tend to use defensive roles, and the opposity when playing againts easy opponents.
One of the most problematic aspects of this work is the simultaneous learning of preferences of all agents, this changes continuously the environment and can affect other agent preferences.
Current work
We plan to measure agent diversity using Hierarchic Social Entropy [2] adapting its approach to measure physical-dynamical diversity in robotic teams. We are going to analyse several properties of the ecosystem (stability, adaptability, performance,....) considering the degree of diversity in this system as we said initially
References
[1] http://www.parc.xerox.com/istl/groups/iea/abstracts/MultiagentDynamics/controllingChaos.html
http://www.parc.xerox.com/spl/groups/dynamics/www/multiagent.html
Last Updated: 13th March 2001
