Path: utzoo!censor!geac!torsqnt!lethe!yunexus!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!cs.utexas.edu!uunet!tut.cis.ohio-state.edu!pt.cs.cmu.edu!o.gp.cs.cmu.edu!andrew.cmu.edu!rr2p+ From: rr2p+@andrew.cmu.edu (Richard Dale Romero) Newsgroups: comp.ai.neural-nets Subject: Re: Neural nets applied to Expert Systems? Message-ID: Date: 8 Jan 91 21:04:43 GMT References: <15385@arisia.Xerox.COM> <2670@bimacs.BITNET>, <17355@brahms.udel.edu> Organization: Carnegie Mellon, Pittsburgh, PA Lines: 41 In-Reply-To: <17355@brahms.udel.edu> Chris Bryden writes: >Actually, I heard a nasty rumor that someone had a method by which >they could pull rules, for an expert system, from a properly trained >neural network. I would appreciate any references that would make >or break this rumor. hmmm... sounds like a small, local lecture i attended for a new visiting professor here at cmu. basically, he took a few rules that described an approximation of a system, (he was using a 3-d surface map,) and created a neural network that conformed to those rules exactly. the rules were like: given x in and y in , z will be in ; given x in , z will be in . he made a couple layers that acted as a classifier, then he made a couple more that applied the rules. the whole thing was done by hand, but he said that the whole thing could be automated, which doesn't seem too far-fetched. anyways, then, he trained the whole thing on the actual surface map using back propagation. because of the original rules that were encoded into the network, the new network was fairly easy to pull *more refined* rules out of. the rules were implicit in the original network architecture. basically, the original rules had used linear classifications, and it was evident, after examining the trained network, that non-linear rules had been necessary to encode the surface map. there were some problems raised with the method, though. it was evident that if your original rules were "bad" or you got them from a poor expert, that the rules that would be developed would not match the original form they were encoded with. in other words, if you introduced several erroneous rules, you may never realize that these rules had been changed into something totally different. for example, suppose you had the classification for several rules, but it ended up that this classification wasn't necessary for any rules. the network would learn to ignore this classification and it would be hard to realize that this was occurring. suffice it to say, that the pulling out of these more refined rules is not fool-proof. sorry i don't have a reference, but i'm not even sure that an entire paper has been completed. -rick