Learning in multi-agent systems

Project Details

Description

The study of this project is the aspect of learning in multi-agent systems. It is extremely difficult to determine at design time the interactions and actions of aal agents in a correct and complete way. Since this would mean that one knows what can be expected from an agent in advance or which agents will be availble and so on. Therefore it is advisavle to incorporate a learning component in the agents, so they can learn how to behave otimally and independent in a changing environment. The traditional learning algorithms in the context of artificial intelligence however cannot be used straightforwardly in multi-agent systems. this for several reasons: the environment is non-stationary and/or there is no central control an/or agents do not have complete perception of their environment and/or the information agents receive is delayed and incomplete. First the technique of reinforcement learning will be adapted for use in multi-agent systems. This technique is a good candidate since it doesn't need an a priori model of the environment and can therefore be used when agents have little knowledge of their domain. Second the domain of Learning Automata will be studied as a candidate for autonomous coordination of agent actions. Finally evolutionry game theory will also be investigated as a possible approximation.
AcronymOZR578
StatusFinished
Effective start/end date1/01/0131/12/03

Flemish discipline codes

  • Information and computing sciences
  • Mathematical sciences

Keywords

  • artificial intelligence
  • machine learning
  • distributed systems
  • game theory