Jörg Denzinger's

AI methods for the development and set-up of computer games: automated generation of NPC character behavior

Many computer games use computer controlled entities (so-called non-player characters or NPCs) that either help or fight the human player. How good or bad and realistic or not NPCs are has a big influence on how much a particular human player likes a particular game. And the quality of NPC behavior is rather subjective, too. But also boredom is a crucial factor, since most human players will figure out how to beat a particular strategy for an NPC over time. In many games, NPCs also have to work together and usually one AI controls several or even all NPCs. While many games offer several such AIs of different strength, each of these AIs usually requires a lot of work to develop and implement, especially if we want to see rather different behaviors of the NPCs.

Our solution approach to these problems (published at IAT 2012 and CIG 2013, see our bibliography page) is to learn behaviors for individual NPCs that include cooperation with other NPCs using communicated intentions. The learning is achieved by playing against one human generated AI and different learning attempts result in selections of behaviors for particular NPC types of different quality. By combining such learned behaviors in a new way each time a human player plays the game, very different overall game play experiences can be created, much more experiences than available in current games, without much human effort.

to our bibliography page on AI and Games.

Last Change: 5/12/2013