Affiliations:
School of Computer Engineering, Iran University of Science and Technology, University Road, Narmak, Tehran, Iran
Correspondence:
[*]
Corresponding author. Akram Beigi, School of Computer Engineering, Iran University of Science and Technology, University Road, Hengam Street, Narmak, Tehran, Iran. Tel.: +98 21 77209020; Fax: +98 21 77240592; E-mail: beigi@iust.ac.ir.
Abstract: Recently multi agent systems are used to solve complex problems. In these systems agents can cooperate when a problem is difficult or impossible to solve for an individual agent. Via learning, the agents attempt to maximize some of their utilities. In multi agent learning an agent learns to interact with other agents and considering their behaviors. By multi task learning, the agent simultaneously learns a set of related problems and with reinforcement learning, an agent learns a proper policy to achieve its goal. In learning process, using the experience of teammate agents by simple interactions among them is very beneficial. In this paper we have presented a simple model of agents’ interactions using operators of an evolutionary algorithm. Applying the proposed model has improved significantly the performance of multi task learning in a nondeterministic and dynamic environment, specifically for the dynamic maze problem. The experimental results indicate our claim.
Keywords: Interactions between agents, multi task reinforcement learning, evolutionary algorithms, dynamic environment
