VoodooExtreme have posted up the first part of their Ask Paul Tozour column, the artificial intelligence programmer at ION Storm: Austin where he's busy working on the Unreal Warfare engined thriller, Thief III. Heres an interesting a bit:
Samuel: Do you think that neural-net style, "learning" AI will ever have a place in game design? Is it feasable to have NPC's learn from the players actions without excessive prompting from coded presets.
Learning techniques definitely have a place in the present and future of AI. Black & White proves beyond any doubt that learning in games is possible and it can be extremely effective if you use it in the right way.
The real question is specifically how the learning happens and how it fits with the underlying game design. Moving away from the narrow, overly-scripted AI systems of the past is an admirable goal, and one which I personally strive for, but I don't think we should view machine learning technologies as the sole alternative, and in many cases, they're not necessarily a good idea.
The key is to make sure that your underlying AI system is solid. The best way to achieve that is usually not through machine learning technologies. Learn how to play the game yourself, and evolve your AIs based on your own understanding of the game and its specific needs.
After you have an AI that already plays the game well, then you can think about ways to improve it, and ask the question of what machine learning techniques might be able to make your AI more challenging and/or entertaining than it already is.