Author: Gunnar Andersson
Date: 13:24:32 09/28/99
Go up one level in this thread
On September 28, 1999 at 09:38:30, Robert Hyatt wrote: > > >I don't think this is the issue. Here are some problems that have to be >overcome: > >1. no "evaluation" can pick a good book move without a deep search to go >with it. Why? Because humans can't pick good book moves in most positions >without a lot of prior searching (done by others) to guide them. Hence the >plethora of books like MCO, ECO, BCO, etc... My mistake - I referred to the program learning opening deviations using deep searches for alternative moves to the ones actually played in the database. For this to work out, the evaluation function must (a) be pretty good, or otherwise a huge search depth is needed (b) have a global interpretation, i.e. be comparable between game stages The simple scheme "find best deviation for all database positions, then negamax entire game tree including deviations" gives extremely good results in the domain of Othello, considerably stronger than the other opening book schemes proposed. > >2. frequency of play is not a good guideline in many cases, because a recent >refutation gets swamped by all the times it was played successfully in the >past. Indeed. Strong humans analyze enough games for this to be less of a factor, I suppose. > >3. opening theory has many outright mistakes that are waiting to get found. >At the highest levels of play, opening preparation is hugely important. As of >now, computers are not good at doing this. Automatically searching for tactical refutations seems possible. A potential danger would of course be that the program doesn't understand the positional compensation the side with less material can have. / Gunnar
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.