Towards Human-Like Bots Using Online Interactive Case-Based Reasoning

The imitation of human playing style has been gaining relevance in both the Artificial Intelligence for Games research community and the Digital Game industry over the last decade, achieving a special importance in recent years. The goal of these virtual players is to deceive real players and be perceived just as another human player. Although this challenge can be addressed using different Imitation Learning techniques, classic supervised learning approaches do not usually work well due to the violation of the independent and identically distributed assumption for random variables. No regret algorithms in online learning settings seem to outperform previous approaches. In this work we describe an interactive and online case-based reasoning system in which the bot gives control to the human player when it reaches game states that are not well represented by cases in its case base, and regains control when the game states are known again. Results show that (1) the amount of human intervention decreases rapidly, (2) the case base needed to achieve reasonable imitation is considerable smaller than that used in a non-interactive approach (3) the resulting agent outperforms other agents using non-interactive CBR.