AI and Kohonen Networks

[JavaSnob77]

Are there any implications for the use of Kohonen nets as an artificial intelligence device?  Every website mentioning them praises them as super-versatile and super-useful, but I'm looking for an AI-based implication.  Any ideas?

Thanks,
Rob

[sonicviz]

Check out http://www.generation5.org/jdk/ , they have a Kohonen implementation.

[JavaSnob77]

Thanks, but I code my own implementations.  I don't like relying on the code of others, not that it may be faulty, but rather I like the sense of pride I feel after I've finished with a challenging project.

Games published by our own members! Check 'em out!

[sonicviz]

snob

[digitprop]

What do you mean by 'implications'? I think Kohonen networks can be quite useful for pattern recognition / categorization, but in the end it would probably not make much difference what statistical analysis method you use (Kohonen or other NNs, or even Markov models, etc.).

The much more intriguing questions, IMHO, is how to present the problem / data structure to the network in order to get a decent performance (in terms of the quality of the categorization / pattern recognition).

Game situations can get very complex, and thus the number of potential patterns can be very high. On the other hand, the number of actually encountered patterns in a gameplay is comparatively low, so it will be difficult to get a decent statistical analysis - unless one is able to reduce the size of the pattern space by transforming the game situations into abstract, low-complexity states.