[ranch vermin]

So Id like you consider a perceptron as an "approximating search engine."...  but in this one, the perceptron is looking inside another model, which is provided by the environment. (excuse me if youve heard that one plenty of times, im trying to do something slightly original here.)

So these perceptrons, as they evolve, hillclimb, or anneal or WHATEVER,  are maximizing a score function, which itself is hard coded, fixed in the system.

When your looking in this "model generated from the environment" the first thing something thinks of is,  "oh im developing an activity, or some reaction for it to do"  but this is actually only thinking about it in a boring obvious way!

These hillclimbing perceptrons could do lots more than just this!

Perhaps you could score generating 3d from 2d!   and after that you could clear up alot of object distinctions and really neaten up your system before you go develop the action.

So theres alot fine tuning that some prior reorganization could help you with,   but it could just be scoring functions the same!

[LOCKSUIT]

So it does backprop with a cost and finds the optimal actions for each given events?

I'm not sure if you should give us more details, or summarize your OP into a smaller point :P Try one or both though!

[8pla.net]

Multi Layer Perceptron

http://aihax.com/MLP/

[ranch vermin]

its a form of "generate and test" isnt it.

Thats why neural networks can have a go at brute forcing ANYTHING,   but in some cases they dont do so well,  (they have problems with a needle in a haystack) but you really get imaginitive with how your going to solve your problem.

Thats all this thread really means,  that when your writing a network,  you can get a neural net and have a go at anything u dont want to write directly!