Multi-Agent Cooperation Using Genetic Network Programming with Automatically Defined Groups

Abstract. In this paper, we propose a genetic network programming (GNP) architecture using a coevolution model called automatically defined groups (ADG). The GNP evolves networks for describing condition-action relations for agents. By applying ADG to GNP, we evolve different networks in order to realize the cooperation of multiple agents with different abilities. Computa-tional experiments on a load transportation problem show that appropriate networks are obtained with taking account of the ability of agents. 1 GNP with ADG Genetic Network Programming (GNP) has been proposed [1] for describing successions of processing. It is inspired from GP, however it does not have a tree architecture, but has a network architecture. Fig. 1 shows an example of networks obtained by the GNP. There are three types of nodes in a network: a start node, judgment nodes (diamonds), and processing nodes (circles). Agents make decisions on their behaviors according to the obtained network. For example, an agent examines whether it carries a load or not in the first judgment node “Carry? ” If it is carrying a load, it moves to a goal according to the processing node “Move to goal. ” Otherwise, it will go to a place with heavy loads. Using network architecture GNP can obtain a succession of actions as shown in Fig. 1 while GP realizes a succession of actions by returning to a root node from a terminal node in which an agent can take an action. When each agent has a different ability, a different network should be obtained according to the ability of the agent. We employ a coevolution model called automatically defined groups (ADG) [2]. While the model was originally proposed for GP, we apply it to GNP in order to develop different networks according to the ability of agents. We show the effectiveness of the GNP with ADG by computational experiments on a load transportation problem [2].