Author: Will Singleton
Date: 23:49:00 06/22/00
Go up one level in this thread
On June 23, 2000 at 02:03:45, Bruce Moreland wrote: >On June 23, 2000 at 01:13:09, Will Singleton wrote: > >>On June 22, 2000 at 21:35:02, Graham Laight wrote: >> >>>When we've discussed NNs in the past, as far as I remember (and excuse me if >>>I've missed some threads where this has already been discussed), we've always >>>discussed them in terms of creating one giant, monolithic NN. >>> >>>In a flash of inspiration last night, it suddenly occured to me that what you >>>actually need is one NN for each type of "positional factor" you are trying to >>>recognise. >>> >>>For example, you could build a NN of an appropriate size, and train it to look >>>out for cases where a "smothered mate" was likely to become possible - using a >>>set of positions where it either is, or isn't likely. Then, for each node that >>>is to be evaluated in a game, you can pass that position to this NN, and come up >>>with an assessment of whether this particular positional weakness (or strength!) >>>exists. >>> >>>When you have got 50,000 such NNs working, then "vous voila!" - you have your >>>player with grandmaster level skill. >> >>>I think that this would be a good use for multiple processor (or multiple >>>computer) systems. And, in reality, wouldn't it be a good model for how the >>>human brain works? >>> >>>In summary, instead of thinking about one big NN, think about having lots of >>>little NNs to play chess with. >>> >>>-g >> >>I think your idea has a lot of merit, but not sure this is a new idea. I also >>think it could take years to design a system to coordinate a set of eval >>functions which could choose the proper neural weight set (or sets). Would be a >>process of tuning the training function, then playing games to judge the result, >>then modifying the training function and generating new weights. >> >>It might be very difficult to get this to work. NN's seem to work best where a >>huge set of games can be used to figure out the training function params, and >>the resulting weights. Lots of trial and error. Chess doesn't lend itself well >>to such training. >> >>Will > >I think there is little substance to this idea, and furthermore I expect that >there is little truth to the notion that this has anything to do with human >thought. > >A series of in-parallel recognizers. Fine. But there is absolutely nothing >about how to relate them, and only one pattern out of 50,000 proposed patterns >is enumerated. > >We have the title for yet another novel, but I would like to see the first >chapter, or at least the first paragraph, before I consider giving Graham a cash >advance. > >bruce Agreed. There's an interesting article at http://www.lucent.com/press/0600/000621.bla.html Not really relative to this particular discussion, but perhaps in the future... Will
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.