Incremental and Mutual Adaptation in Multiagent Systems

There are two streams of research when combining multiagent systems and learning. One regards multiagent systems in which the agents have learning capabilities and they learn from the environment in which they operate. The feedback the agent receives from the environment includes knowledge about the other agents in the system. The second direction investigates the issue of multiagent learning where the focus is on the interactions among the learning agents. Each agent learns directly from the other agents that operate in the same environment. In this work, we investigate how agents can learn strategic behavior in a teacher-learner model. The notion of the teacher here should be understood as a \trainer." In certain research on automated learning, a \teacher" is presupposed to be benevolent with his student. In a general multiagent learning system, however, we might have benevolent agents as well as manipulative agents. A manipulative agent, acting as a trainer, will instruct the learner in a way that causes it to act in the future for the trainer's beneet. Each agent computes a function that expresses the agent's satisfaction level. This function depends on two terms. One represents the learner's gain from the action it has performed, and the state to which the world has moved. The second term represents the degree of approval of the trainer according to the learner's behavior so far. The general model enables an agent to learn from many trainers, considering the weighted degree of approval of its trainers in its environment. Each agent adjusts its behavior to increase its satisfaction level in order to adapt to one another. Our aim is that the agents learn to coordinate their behaviors, i.e., the agents will learn to adapt themselves. We present preliminary results of experiments we have conducted. We also propose an algorithm for incremental adaptation in a multiagent system, based on the trainer-learner model where the agents behave in order to maximize their satisfaction levels.