Author: Matthew Hull
Date: 12:40:17 05/12/05
Go up one level in this thread
On May 12, 2005 at 14:50:41, Michael Yee wrote: >On May 12, 2005 at 12:58:00, Matthew Hull wrote: >>On May 12, 2005 at 12:31:05, Steven Edwards wrote: >>>On May 12, 2005 at 12:18:10, Matthew Hull wrote: >>>>On May 12, 2005 at 12:09:11, Steven Edwards wrote: >>>>>On May 12, 2005 at 11:54:16, Matthew Hull wrote: >>> >>>>So, the initial test shows some success. You also indicate in another post that >>>>a more comprehensive problem set will be the next step, at which point you will >>>>have a mate-attack organism thingy with which to work. How many other organisms >>>>do you estimate will be needed to get a basic cognitive process to start taking >>>>over from the toolkit? Do you have a preliminary list of these in mind? >>> >>>I have it working on the 1,001 position suite BWTC. Later, I'll try automated >>>construction of a multithousand position suite from appropriate PGN data. >>> >>>Symbolic has a knowledge sequencer named KsSurveyor whose job is to identify >>>strategic themes in a position. Each recognized theme is posted as a Theme >>>instance in the Instance Database in the position search tree node for that >>>position which is later used by the rest of the planning process. My idea at >>>the moment is to have (at least) one species/organism for each theme. This >>>organism will be used to determine theme matching, any moves that will promote >>>the theme (along with ranking), and (via a one ply search) help determine >>>countermoves that work against the theme. >>> >>>The list of themes will be stolen from various chess books I have. Some themes >>>like MateAttack are simple in that they aren't parameterized. Other themes will >>>have target, region, and sequencing parameters. >> >> >>I see this process as similar to NN technology in terms of the utilization of >>processing resources. The bulk of the compute resource are consumed ahead of >>time, and the generalized "understanding" stored for later reference. The >>application of these learned things will require further effort in a game, but >>most of the effort will have already been computed and stored in generalized >>form. A traditional AB searcher remembers almost nothing and must re-compute >>it's "knowledge" frequently and afterward remembers nothing, except what may be >>in the transient hash table. The exception is the EGTB, whose compute resources >>were already consumed and don't need recalculation. The EGTB is the only thing >>"remembered" between games, except some basic position/book learning. But even >>the hash table and EGTB are not really "understanding", but just rote memeory. >> >>The conservation of computing effort is what I like most about your approach, >>and other approaches like NN. > >The observation about storing knowledge/generalization ability is nice. But I >think there are some key differences between neural networks and genetic >algorithms. > >NNs are merely a class of functions/models with parameters. (Well, their >structure happens to be able to construct new features from more basic ones, >which is neat.) The parameters can be fit/learned using backpropagation or any >other optimization technique. > >A GA is merely a pretty robust optimization technique that can be used to fit >the parameters of a model (e.g., a NN), or, in general, optimize any objective >function. > >In Steven's case (I think), he chose subsets of microfeatures (templates) as his >family of models and used a GA to optimize an objective function (test suite >performance) over possible models. A key difference between his family of models >and NNs is that NNs are hierarchical. I see it mainly as a re-distribution and conservation of computational resources. The AB searcher does the bulk of its computation at the point of application. The NN and/or GA-cognitive-function do the bulk of their computation before the game is ever played. The CPU consumed at the time of application is dwarfed by the CPU expended in the "learning" process. Also, the AB computational results are thrown away almost immediately, while the other approaches attempt to preserve hard-won learning results for future efficient use, thus conserving effort and hopefully contributing to dynamic "growth". Some AB searchers are using offline processing to find optimal parameter settings, but once the optimum has been found for that design, the "learning" is over. It seems like Steven's approach is an attempt at open-ended learning capability, which an interesting goal. > >Michael
This page took 0.01 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.