Co-evolutionary agent model for adaptive behavior

How can a system be more adaptive and efficient? In a multi-agent system (MAS), it may be a good idea to evolve each agent separately and evaluate them together in the common task. We propose a multi-agent system composed of adaptive agents, which are incorporated in MAS environments to pursue their goals separately and then co-evolved together in order to make them more adaptive and efficient. In this system, each agent is embedded with an inner-learning unit (LU), which concentrates on the reinforcement algorithm with historical local information and an external co-evolutionary learning unit that acquires global reinforcing information from the environment and other agents. Agents adjust action strategies according to the evaluation of global rewards. Through such operation, agents are expected to co-evolve together to achieve a global optimized result. To store the best result of MAS ever gained in the learning process, a shared memory unit is used. Compared to the widely used "top-down" method, this approach emphasizes a co-evolutionary method about distributive control, which aims at unifying the individual's self-evolving ability and the system's global information. To demonstrate the effectiveness and efficiency of this approach, the predator/prey domain is used as an example of simulation in which agents represent different predators and prey. The result from the simulation shows that the proposed approach has a high potential for distributive co-operative problem.