Multi-agent systems which include one or more learning agents share some of the concerns of autonomous agent learning. For example, the issue of when to reason, or when to act is even more pertinent in this context. There are important distinctions between the two approaches. Multi-agent learning offers radically different solution to some of the problems in learning. A robot can become more correct (or more reactive) not only by learning from experience, but by communicating with other agents (artificial or human agents). No wonder that the attention of several researchers working in this area has turned to various architectural issues, all of which have something to do with communication. The design should determine when, how and with what purpose should the individual agents communicate. Various systems differ in how they approach these questions. Basically, the learning agents can communicate:
- before the learning / problem solving phase,
- during the problem solving phase,
- before the problem solving phase, but after the individual learning phase,
- during the individual learning phase.
Expressed differently, the agents can be involved in distributed data gathering, distributed problem solving or distributed learning. Of course various hybrid solutions may exist too.