Author: Michael Yee
Date: 08:05:46 05/12/05
Go up one level in this thread
On May 12, 2005 at 10:50:41, Wayne Lowrance wrote: >On May 12, 2005 at 09:06:29, Michael Yee wrote: > >>On May 12, 2005 at 07:00:35, Peter Fendrich wrote: >> >>>On May 12, 2005 at 01:57:19, Steven Edwards wrote: >>> >>>>Symbolic: Status report 2005.05.12 >>>> >>>>The preliminary version of the genetic algorithm framework is complete. >>>>Complete details would take many pages, so I'll just post a brief overview of >>>>the initial experiment and the results. >>>> >>>>The shortest mating test suite, Bloss (fourteen positions), was chosen for the >>>>first tests because of its brevity. For detecting mating attack moves, a >>>>species template containing twenty microfeature recognizers was defined. A >>>>habitat containing one hundred randomly generated organisms of the species was >>>>generated with the organisms' selective power measured against the best moves in >>>>the Bloss suite. The highest ranking initial organism did fairly well, >>>>correctly selecting the best move in eight out of the fourteen problems. >>>> >>>>After the initial habitat generation, the habitat is repeatedly cycled. Each >>>>cycle consists of picking two parents (with a selection bias based on merit), >>>>producing an offspring, mutating the offspring slightly, measuring the >>>>offspring's merit against the suite, and then inserting the offspring into the >>>>habitat (if it's better than the least fit occupant; the least fit occupant is >>>>removed). A new offspring organism that outranks all the earlier ones is >>>>displayed on the ChessLisp console. >>>> >>>>After 101 cycles, a new champion organism was produced that matched nine of the >>>>Bloss problems. After 285 cycles a ten matcher was found. And on cycle 411 an >>>>organism was found that matched eleven. A twelve of fourteen matcher was >>>>produced on cycle 453, and a thirteen match organism appeared on habitat cycle >>>>1297. >>>> >>>>More to come. >>> >>>Hi Steven, >>>It would be interesting to hear more about this GA approach. >>>Maybe you have a hompage and can put som information there? >>>/Peter >> >>I second Peter's request :) >> >>Would it be possible to explain the nature of the features and how you computed >>the fitness? For example, were the microfeatures pre-constructed and the >>chromosome's genes the weights? Or were the microfeatures themselves the genes >>of the chromosome and was the fitness computed by having the chromosome score >>each move according to some function of the presence of the chromosome's >>particular features? >> >>Thanks, >>Mich > >huh ? we talking chess here ? >Wayne Sort of... (although my post does sound like double-talk now that I'm re-reading it!) 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? 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.