Cooperative Behavior Rule Acquisition for Multi-Agent Systems by Machine Learning

Abstract

Multi-agent systems are the systems in which several interacting, intelligent agents pursue some set of goals or perform some set of tasks (Wooldridge, 2000). In multi-agents systems, each agent must behave independently according to its states and environments, and, if necessary, must cooperate with other agents in order to perform a given task. Multi-agent systems have more robustness and flexibility than conventional centralized management one. However, it is difficult to beforehand predict the action of the agent and give the action rule for the multi-agent systems, because the interaction between agents is complicated. The acquisition of an autonomous agent’s intellectual action rule is a very interesting topic. Recently, extensive research has been done on models such as Robocap (Stone & eloso, 1996, Matsubara et al., 1998). Studying computation models of cooperative structure to accomplish a given task is known to be a difficult problem (Jeong & Lee, 1997).In the field of self-learning reactive systems, it is not even desirable to have a clear idea of a computational model. Thus, being adaptable, autonomous agents imply minimally pre-programmed systems. Numerous studies regarding autonomous agents in the multi-agent systems have been conducted. Nolfi and Foreano (Nolfi & Floreano, 1998) simulated a pursuit system with two agents (predator and prey) in real environments. They evolved both agents reciprocally using a genetic algorithm. Jim and Lee (Jim & Lee, 2000) evolved the autonomous agents with a genetic algorithm. Zhou (Zhou, 2002) used both a fuzzy controller and a genetic algorithm. The fuzzy function displayed the position of the agent, and the genetic algorithm was used for learning. Fujita and Matsuo (Fujita & Matsuo, 2005) learned the autonomous agent using reinforcement learning. The reinforcement learning method involves developing an agent’s behavior by means of the interrelationship with the environment and resulting reinforcement signals. The reinforcement learning method can guarantee learning and adaptability without precise pre-knowledge about the environment. In this chapter, we focused on the problem of “trash collection”, in which multiple agents collect all trash as quickly as possible. The goal is for multiple agents to learn to accomplish a task by interacting with the environment and acquiring cooperative behavior rules. Therefore, for a multi-agent system, we discuss how to acquire the rules of cooperative action to solve problems effectively. First we used a genetic algorithm as a method of acquiring the rules for an agent. Individual coding (definition of the rule) methods are performed, and the learning efficiency is evaluated.