A generic approach for generating interesting interactive Pac-Man opponents

Abstract

This paper follows on from our previous work focused on formulating an efficient generic measure of user's satisfaction (`interest') when playing predator /prey games. Viewing the game from the predators' (i.e. opponents') perspective, a robust on-line neuroevolution learning mechanism has been presented capable of increasing --- independently of the initial behavior and playing strategy --- the well known Pac-Man game's interest as well as keeping that interest at high levels while the game is being played. This mechanism has also demonstrated high adaptability to changing PacMan playing strategies in a relatively simple playing stage. In the work presented here, we attempt to test the on-line learning mechanism over more complex stages and to explore the relation between the interest measure and the topology of the stage. Results show that the interest measure proposed is independent of the stage's complexity and topology, which demonstrates the approach 's generality for this game.