Evolutionary Strategies of Intelligent Agent Training

Abstract

Groups of interacting agents are able to solve complex tasks in a dynamic environment. Robots in a group can have a simpler device than single stand-alone robots. Each agent in the group has the ability to accumulate interaction experience with the environment and share it with other members of the group. In many cases, the group behavior is not deduced from any properties of its parts. The paper proposes an approach to modeling the mobile agent group behavior that is busy with a common goal. The main purpose of the agents is to study the greatest territory at minimal time. The agents interact with the environment. A control of each agent is carried out by a modified neural network with restrictions imposed on it. Weights of the neural network are chosen by a genetic evolution method. The agents compete among themselves for obtaining the greatest reward from the environment. An efficiency of the proposed model is confirmed by some convergence speed tests by computer simulation. The proposed model can be applied to a group of robots that perform search tasks in a real physical space.