Computer Games, a boon or a bane
... are lots of trade-o_s to be considered when striving for these features. Trade-o_s between speed and memory are well 1 known, but the same applies to other features. Currently, nearly all AI research focuses exclusively on computational e_- ciency, neglecting other goal features. Because of the trade-o_s involved in multi-objective tasks, this leads to the somewhat counter-intuitive e_ect that { although focusing on something obviously important like computation speed { research steers away from possible applications, further widening the gap be- tween research and application. What is the point in specializing an algorithm for solving a machine-scheduling problem 75% faster if the result is an extremely specialized version that cannot be reasonably adapted anymore to handle dy- namic changes, like a machine breakdown or the change of an order? Oh yes, I forgot, you can publish a paper that will surely be cheerfully praised by the academic community because you beat some abstract toy-problem benchmark... It is much easier to compare approaches and results with respect to only one feature, which may be a reason why everyone seems to bow to the deity of computational e_ciency. It is certainly correct that in general, di_erent applications need di_erent feature sets to be supported, which makes a direct comparison somewhat di_cult. Looking at di_erent feature sets, however, does not mean that we enter the realm of arbitrary interpretation. Richer metrics to compare and guide solution development must be developed, and respective benchmarks be established that consider richer feature dimensions. Given our human { real-time, dynamic and only partially known { life environment, it is hard to believe that a niche area for systems that neglect most of these features would receive much attention in this extended feature view. You may now see why I do not like games like chess as a research object. Solving chess in one year from now instead of three may be good for political reasons. The scienti_c contribution, however, is hardly anything notable because the respective methods for this boring one-feature-focused problem are already on such a specialized level that an application for anything else is practically ruled out. Real" computer games, on the other hand, make it necessary to keep a great number of features in focus. With their mind-boggling complexity of large-scale simulations with realistic physics and vast interaction possibilities, they can easily be seen as an equivalent to the real world with simpli_ed sensing and actuator procedures. In contrast to the real world, experiments can easily be controlled and repeated. There are of course also several research groups working in the direction of considering a larger number of relevant features. Initiatives like RoboCup (which I do not consider to be a computer game; but nevertheless, it has some common features) have spurred this development. The same is true for the robotics area, and more generally, in areas where people have to contribute to a real application and things have to actually work. Unfortunately, this application focus often turns out to be only yet an- other case study", and results are rarely analyzed from a more general point of view. An application like computer games is no di_erent in this respect, and the right balance between application-speci_c hacking" and general technol- ogy/approaches is not easily achieved. It is important to keep this in mind - 2 especially in an exciting application area like computer games, where lots of tremendously impressive technology blends together with your own. From look- ing over paper submissions and presentations in this area, I have to say that many supervisors seem to let their students loose on this great looking ...