Learning believable game agents using sensor noise and action histogram

A believable game agent can help players to immerse in the game world and maintain their suspension of disbelief, thereby making the game more enjoyable and satisfying. This paper explores the use of two main ideas to acquire believable reactionary behaviours in game agents through learning low level human tendencies in depressing input keys. First, sensor noise is introduced to simulate errors in human judgment and its parameters are optimized through the evolution of the game agent. Second, indirect modeling of human reactionary behaviour tendencies is achieved by using input action histograms as optimization objectives. Two types of histograms are explored, the action histogram and the action sequence histogram. The resulting game agents with evolved sensor noise are able to improve its believability without compromising gaming performance in training, and demonstrate good generalization capability on four other previously unseen test tracks. In a user study involving 58 respondents, the same game agents are also evaluated to be more believable compared to one evolved for performance alone.