Evolving the Behavior of Autonomous Agents in Strategic Combat Scenarios via SARSA Reinforcement Learning

Computational agents of commercial Real Time Strategic (RTS) games mostly have their behaviors designed via simple ad hoc and static techniques, which require manual definition of actions. Thus, such agents are not able to adapt themselves to diverse situations and their behavior becomes predictable along the game, enabling human players to eventually discover the strategies used by them. This work proposes a modeling approach for the use of SARSA reinforcement learning technique applied to combat situations in RTS games. This technique enables that computational agents evolve their combat behavior according to actions of opponents. The performance of this technique was evaluated using a Starcraft based simulator. The experiments showed that agents were able to improve their behavior, developing knowledge to decide about the best actions during different game states and using this knowledge in an efficient way to obtain better results in later battles.