Monday, May 11, 2015

[dqujxtyw] Encoding human chess knowledge

Assume there still exists some chess wisdom within human players that can improve chess programs.  How can we usefully extract and encode that knowledge?

One vague idea: In a database of human games, identify via computation human moves that are objectively good but take a program an unusually long time to discover.  "Objectively good" is approximated by long computation, perhaps Monte Carlo to discover strategically good moves.  Then, perhaps working with the human who made the unusual moves, ask to describe the reasoning behind them, some of which can hopefully be formulated into an algorithm that allows the computer program to discover the move quicker (while not decreasing the overall level of play).  This assumes that anything a human can quickly and accurately calculate, a computer can also.  This assumption may be false for pattern recognition AI tasks, which chess evaluation is, because of radically different architectures between the brain and electronic computers.

The motivation was, if a person wanted to maximize his or her impact or legacy on future chess in terms of increasing the general level of play, it might be by improving an open-source chess program.  In the past, it may have been to write a book, but programs encode knowledge more precisely than prose.

No comments :