Improving Temporal Difference game agent control using a dynamic exploration during control learning

摘要: This paper investigates the use of a dynamically generated exploration rate when using reinforcement learning-based game agent controller within dynamic digital environment. Temporal Difference learning has been employed for real-time gereration reactive behaviors variation classic arcade Pac-Man. Due to nature environment initial experiments made static, low value utilized by action selection during learning. However, further were conducted which prior genetic algorithm. Results obtained have shown that an improvement in overall performance may be achieved is used. In particular, if algorithm controlled measure current agent, gains achieved.