MACC'97 Session: Learning, Evolutionary Computing
This paper proposes and evaluates a reinforcement learning system, L-ANA (learning agent network architecture), in which a behavior-selection network using spreading activation, such as ANA is applied to exploitation-oriented reinforcement learning. L-ANA is more robust and flexible in the face of dynamic environmental changes than conventional exploitation-oriented reinforcement learning systems such as profit-sharing, because spreading activation is used to assign reinforcement values. Thus, L-ANA is well suited for autonomous systems that operate in complex dynamic environments such as the real-world and the Internet. In L-ANA, memory agents that are autonomous entities similar to the agents in swarm-made architectures, control learning autonomously and the spreading activation is implemented through their coordination. In addition, by manipulating the parameters of the spreading activation, it is easy to tune the learning characteristics. We verified the fundamental effectiveness of L-ANA by simulation.
reinforcement learning, memory agent, spreading activation, robustness, real-world
PS
file(+gzip) (in Japanese)