Author: Michael Yee
Date: 11:36:22 05/12/05
Go up one level in this thread
On May 12, 2005 at 12:04:45, Steven Edwards wrote: >On May 12, 2005 at 11:05:46, Michael Yee wrote: > >>The gist of my question was essentially : what were the *chess* features like? >> >>For example, did each chromosome somehow select/"design" its own >>features/patterns from some feature space: >> >>- feature1 : >= 1 slider piece in same file as opponent king >>- feature2 : <= 2 flight squares for opponent king >>- feature3 : >= 2 pieces attacking h7 >>- etc. >> >>Or did each chromosome just have weights for some pre-defined features (similar >>to a linear static evaluation function)? >> >>- kingDefenseWeight = 0.8 >>- pawnShieldWeight = 0.5 >>- etc. >> >>Option 2 is unlikely since you could fit those without using GAs. So my main >>question was : what was the space of possible features? > >A microfeature is a easily recognized pattern without instance variables. >Currently, thery are hardcoded but this is subject to change. A species is a >template made from a subet of the available microfeatures. (Only the MateAttack >species is being developed at the moment.) Here is the microfeature list used >in the first experiment: > > >(defconstant MateAttackMfSymbolVec > (make-vector-by-content > '( > MfAdjCheck > MfAdjDefenderCapture > MfAdjDefenderDecoy > MfCapture > MfCheck > MfDiscoveredCheck > MfDoesCheckmate > MfDoesMateIn2 > MfDoubleCheck > MfFlightDecrease > MfFlightIncrease > MfForceCEInterposition > MfForceCEKingNoncapture > MfNoCheck > MfResponseCountEq1 > MfResponseCountEq2 > MfResponseCountGt4 > MfResponseCountLe4 > MfThreatenMateIn1 > MfThreatenMateIn2 > ))) > >Your option two is fairly close to describing the initial experiment. However, >I'm working on the automated synthesis of new genes produced from conjunctive >normal form expressions of existing ones. > >The merit calculation in the initial experiment is simple: it's the total best >move match count of the training test suite. The next revision here will be to >measure the average raw weight total distances between rankings of best moves >and non-best move groups summed over all the positions. This is needed for >several tastsk, one of them being the ability to understand thresholds (how many >mating attack moves exist, and is a mating attack justified at the current >exploration effort level). Thanks for the explanation. Your "templates" sound like subset selection, which is certainly a harder problem then just learning weights. Overall, your approaches sound very interesting and seem to go beyond what's been done before. For example, move ordering has been studied a lot, but I don't think it's been studied in a situation/plan-specific way (like move ordering/selection during a mating attack plan). Michael
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.