Author: Ed Schröder
Date: 09:47:44 11/14/03
Go up one level in this thread
On November 14, 2003 at 11:44:03, Georg v. Zimmermann wrote: >Hi Ed, > >this is very interesting information, and good ideas I think. > >I wonder, did you ever test your learning algorithm against that of other >programs ? > >This is harder to accomplish of course than testing the engine as a whole. > >What first comes to mind is something like this: play 1000 (at least !) fast >games against another program, with learning turned off. Then play another 1000 >with learning on. > >Then compare ratios, if program A loses 300-700 in the first test , but loses >only 450-550 in the second, it has much superior learning. > >Now the problem might be that this tells us which program has had a better book >to start with, because the books need to be rather large in the first test to >avoid repeated games. Maybe one should just not count repeated games in the >first test. > >Have you done something like that ? Hi Georg, Learning is a nasty thing to test. What I have done to test learning stuff is self-play matches. The advantage of self-play matches is that you have 2 opponents of equal strength (and equal books if you are testing book-learning) and further excluding all randomness such as playing only from pre-defined positions and with reserved colors. In such a test environment playing only 100 x 5/40 games is more than sufficient to proof if a learning algorithm is working or not. My best, Ed
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.