krulwich@ils.nwu.edu (Bruce Krulwich) (04/13/91)
In article <pluto.670117666@zaius>, pluto@zaius (Mark Plutowski) writes: >mamisra@pollux.usc.edu (Manavendra Misra) writes: >>I was wondering if there was research going into developing "neural" >>algorithms for playing games like Chess. [...] > >A quite famous (and successful) neural-network game player is the NeuroGammon >implementation of backgammon. (Look in 1989 or 1990 NIPs proceedings. I am >unsure about the author's name but Tesauro (sp?) comes to mind.) The >implementation was most interesting to me due to the way it was trained. >(That's right, trained, not programmed, as was Deep Thought.) This isn't quite true (at least not in the version of NeuroGammon that Tesauro was giving talks on about 2 years ago). If you look at the way NeuroGammon was trained, it _started_ _out_ knowing what tactical concepts were important, such as lines with exactly two pieces on them. (I don't remember others, but there were a few more.) These concepts are what enabled it to learn, and they were _wired_ _in_. Yep, that's right, the learning didn't do the feature selection, the programmer did. Oh, sorry, not "programmer," rather "designer." Much different. This is why the system did almost as well learning without any hidden units. Usually hidden units encode learned concepts, but here there were none. All the key concepts which are used in playing were encoded in the inputs. Don't get me wrong -- NeuroGammon is a good piece of network engineering, and it's very interesting that the network (without any projection) was able to play expert-level backgammon. I'm merely objecting to the claim that it learned the concepts necessary to play. If Tesauro did a later version than I saw which did this, I retract this message, but everything I've seen has the features preselected, which avoids all the hard issues in machine learning. Bruce
blenko-tom@cs.yale.edu (Tom Blenko) (04/16/91)
In article <1356@anaxagoras.ils.nwu.edu> krulwich@ils.nwu.edu (Bruce Krulwich) writes: |Don't get me wrong -- NeuroGammon is a good piece of network engineering, and |it's very interesting that the network (without any projection) was able to |play expert-level backgammon. I'm merely objecting to the claim that it |learned the concepts necessary to play. If Tesauro did a later version than I |saw which did this, I retract this message, but everything I've seen has the |features preselected, which avoids all the hard issues in machine learning. Perhaps you're just speaking casually here, but that seems inappropriate when you are mounting such a criticism. Certainly some of the features are pre-selected (or "built-in"). Is it possible to have a system in which this isn't true? I think not. Since each element of this system learns what to "do next", given some history of what has transpired previously, and it IS able to win backgammon matches, why ever wouldn't one say that it has learned the concepts necessary to play? Tom