Mutually Supervised Learning in Multiagent Systems 

Claudia V. Goldman and Jeffrey S. Rosenschein 
Computer Science Department
Hebrew University
Givat Ram, Jerusalem, Israel
ph: 011-972-2-658-5353
fax: 011-972-2-658-5439
email: clag@cs.huji.ac.il, jeff@cs.huji.ac.il


Abstract:

Learning in a multiagent environment can help agents improve their 
performance. Agents, in meeting with others, can learn about the 
partner's knowledge and strategic behavior. Agents that operate in 
dynamic environments could react to unexpected events by 
generalizing what they have learned during a training stage.

In this paper, we propose several learning rules for agents in a
multiagent environment. Each agent acts as the teacher of its
partner. The agents are trained by receiving examples from a sample
space; they then go through a generalization step during which they
have to apply the concept they have learned from their instructor.

Agents that learn from each other can sometimes avoid repeatedly
coordinating their actions from scratch for similar problems.  They
will sometimes be able to avoid communication at run-time, by using
learned coordination concepts.