Portal de Periódicos da CAPES

Competitive Reinforcement Learning in Atari Games

2017; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-319-63004-5_2

1611-3349

Mark McKenzie, Peter Loxley, William Billingsley, Sebastien Wong,

Artificial Intelligence in Games

This research describes a study into the ability of a state of the art reinforcement learning algorithm to learn to perform multiple tasks. We demonstrate that the limitation of learning to performing two tasks can be mitigated with a competitive training method. We show that this approach results in improved generalization of the system when performing unforeseen tasks. The learning agent assessed is an altered version of the DeepMind deep Q–learner network (DQN), which has been demonstrated to outperform human players for a number of Atari 2600 games. The key findings of this paper is that there were significant degradations in performance when learning more than one game, and how this varies depends on both similarity and the comparative complexity of the two games.