Hybrid learning approach based on adaptive resonance theory and reinforcement learning for computer generated agents.

摘要

A new hybrid learning methodology for low-level Computer Generated Forces (CGF) decision processes is proposed. The methodology is implemented combining a recent version of the adaptive resonance theory called ARTMAP-IC and a reinforcement learning technique called the temporal difference (TD) method. The proposed approach adds the ability of automated learning to the ARTMAP-IC, improves the baseline performance and the learning speed of the reinforcement learning, and enables the autonomous agents to learn automatically from unexpected environments with a sufficient level of intelligence and at a low computational cost. The approach proposed here has been tested and verified in a tank battle simulation based on “TankSoar”. Three learning algorithms, ARTMAP-IC, reinforcement learning, and our hybrid learning, were applied to the agent decision process and compared in the simulation. The proposed hybrid learning adapted faster to the environment than the other learning architectures did. The hybrid approach also prevented the expansion of the computational cost for making decisions.