MACC'97 Session: Probabilistic Model in Multiagent Systems
In dynamically changing and unpredictable environment, coming to an agreement for a multi-agent system without explicit communication is an important issue. We will discuss an advantage of probabilistic framework for coming to an agreement. Especially, complex agents that behave according to particular situations, can be modeled by Bayesian (belief) network. In this study, a process to achieve an agreement is formulated as learning of the probabilistic models. Learning algorithm can be derived from probabilistic hill-climb methods. Agents can evaluate a degree of consistency as frequency in learning process. Finally, learning agents find adequate parameters to go along with other agents, then they reach the agreement in the environment. We will show results of learning sigmoidal belief network in several cases. Convenient techniques based on probabilistic framework are also shown.
Probabilistic model, Learning, Focal point, Bayesian network, Cooperative learning
PS
file(+gzip) (in Japanese)