A NEURAL NETWORK-LIKE CRITIC FOR REINFORCEMENT LEARNING

An adaptive agent that contains a reactive network and a critic that supervises that reactive network have been studied. Agent actions are generated in response to stimuli through the reactive network and they influence the ambient environment The critic has a new learning algorithm that recursively enhances reinforcement signals from fixed reinforcement signals by interacting with the environment. The reactive network learns appropriate stimulus-action relations by reinforcement learning. Computer simulation demonstrates that this neural critic is effective in environments where the concepts are embedded in a maze structure. We also suggest similarities between this critic model and the neural circuit in the human brain.