井上 寛康

We focus on a multi-Agent learning where plural agents acquire different specialties to achieve the system goal. This allows the system to solve the deadlock or malfunction problem where agents cannot realize the system goal due to a lack of coordination among the sub-goals pursued by the agents. To this end, this paper proposes an algorithm based on the learning classifier system that divides the sub-Tasks that agents specialize in. Through experiments, it is shown that agents with the algorithm have greater potential compared to agents using the conventional learning classifier system when there are only a few agents in the system or the environment is too large for the conventional learning classifier system to learn effectively.

This paper extends the learning classifier system. (LCS) to introduce the mechanism for recognizing a current situation by determining a boundary between self and others, and investigates its effectiveness in an interaction with an agent. Intensive simulations for adapting an interacting agent by acquiring its internal model have revealed the following implications: (1) the proposed mechanism. gives a higher adaptation to the integrating agent than a random mechanism, the conventional LCS, and previously proposed mechanisms (2) the proposed mechanism keeps its effectiveness, even in a complex internal model of an. agent and (3) the proposed mechanism has the potential to provide autonomy in terms of the precise recognition of the current situation.