Author: Dann Corbit
Date: 18:26:54 02/05/02
Go up one level in this thread
On February 05, 2002 at 18:21:43, Joseph Merolle wrote: >Can you tell me how to axis Positional learning , I know about book learning but >what is this about positional learning ? Any help would be nice guyes. I don't know how the individual authors perform learning. However, there are several ways to go about it. One thing you can do (the simplest possible) is to save a database of positions where your engine screwed up. Typically, you make some move that looks good and suddenly the eval drops into a hole when the opponent does something unexpected. For these, you just store the new eval, and the bad score and the move to avoid. Another thing you can do is TD-Lambda learning. Most programs wait until they finish a game that they lost and then scan over the moves to see where they screwed up. A learning function does a lot of math to change internal weights used by the program. These weights then get written out to disk. Another possibility is to save problem positions into a file. Then, when the engine would be otherwise idle (say, at night) it can pound the stuffings out of them, tracing fowards and backwards along prospective pv's. Eventually, it can find a keen way out of the situation (perhaps by a better move, perhaps by simply tagging the position as dead lost). You could have much more sophisticated learning methods. You could have neural nets. You could have a database with tons of statistics like "What was the time control?" "What was the strength of the opponent that I lost to?" etc. One big problem with learning is that often the learning is completely wrong. Let's imagine two engines playing a blitz game. At some point, the eval drops and your engine loses. So it tags that position. But really, if you had been playing 40/2 then it would have seen that actually instead of being lost, the position is a clear win! So the very position that you have tagged as bad really is the best possible place to be. As you can well imagine, it is not an exact science yet.
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.