Evolving Q-learners for stochastic games: Study on video game agent controllers

The design of the control mechanisms for the agents of modern video games is one of the main tasks involved in the game design process. This task grows in complexity as either the number of different game agents or the number of possible actions increase. An alternative mechanism to avoid hard-coding of agent controllers is the use of learning mechanisms to construct these controllers. These learning mechanisms can be applied during a development phase of the game, making agents learn the best sequence of actions under different circumstances. One of the learning techniques commonly used is reinforcement learning and in particular Q-learners. This paper presents an alternative method to combine genetic algorithms and Q-Learners for stochastic games. The results, using a video game scenario, show that the combination of learning and evolving Q-learner matrices outperforms the sole learning mechanism, independently of the number of learning evaluations.