Author: Rémi Coulom
Date: 03:24:45 10/22/01
Go up one level in this thread
On October 21, 2001 at 07:06:41, James Swafford wrote: >On October 21, 2001 at 02:37:09, Rémi Coulom wrote: > >>On October 21, 2001 at 00:11:23, James Swafford wrote: >> >>If your evaluation function is, say, f(w_1, w_2, ..., w_n), then the partial >>derivative of f with respect to weight w_i is the limit of >>(f(w_1, ..., w_i + epsilon, ..., wn) - f(w_1, ..., w_i, ..., w_n)) / epsilon >>when epsilon goes to zero. In practice, you can estimate this value by measuring >>the ratio above with a very small value of epsilon. > > >That is starting to make sense, but I'll have to sit down and think >about it later today. What is a reasonable value for epsilon? > If your function and weights are integers, this is a not easy to make a good choice. Maybe epsilon = 1 or epsilon = 2 would work. The trick is that epsilon should be large enough so that changing w_i to w_i + epsilon changes the evaluation function if the derivative is not zero. I have not implemented TD(lambda) for my own chess program yet, but I suppose I would make a floating point evaluation function in that case. With a floating point evaluation, things are much easier. For instance, taking epsilon to be 1/1000 of the typical weight value should work nicely. An optimal value for epsilon could be found, but this really is no big deal. Accuracy is not important at all. Remi Remi
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.