giovanni wrote:Hi Rodolfo. Thanks for this interesting thread. I remember that you were working a lot on learning with the Baron program. Would you care to compare these features with the ones in Stockfish-DD-PA_GTB? Thanks in advance.
Giovanni
Hi Giovanni.
I started making experiments with the old Knight Cap by Andrew Tridgell. It was kind of positional learning, as changing value of squares, relative values of pieces, etc. It could never work and, in fact, it was useless.
Some years later, I worked with Romichess by Michael Sherwin. That learning was a simple bonus/malus scoring and book learning, but I enjoyed it a lot, to a point I build an opening book for Romi. It was a learning oriented book, strongly pruned. Only one move per position, while trying to include every suitable opponent reply. My theory was (and still is) that if number of opening lines were reduced enough, then we could improve engine performance by learning. It didn't work at these times, because:
a) Performance level changes in relation of TCs
b) Different opponent books drive the learning towards too many directions
c) A bonus/malus system is a blunder-related system, and it's not true learning if it's based on possible weak moves on ending stage.
Then, I became tester of The Baron by Richard Pijl. Baron learning was a simple position learning with some tricky propagation features. Here, I try to interprete ideas Richard coded:
1) Own position score derives from score after opponent best reply (2 plies propagation)
2) Opponent position score is own score after he made his best move (1 ply propagation - for giving scores to opponent moves too)
Case 1 is possible if own score drops, so we can assume our score was bad evaluated. It's also possible when opponent reply was the ponder move, even if score doesn't drop.
Case 2 is always possible.
In fact, all of that is related to engine strenght and to a correct evaluation. The stronger the engine, the better the learning.
I gave a mere look at Stockfish PA GTB so far. It looks similar to Critter session file, just a bit more sophisticated.
Score backward propagation happens by hash hits. If engine plays several times the same position it can hit previously scored positions while on search, and score changes as a consequence. Engine could then find a better move. Furthermore, score propagation is correctly depth-related. That means a score with depth 20 cannot back-propagate to previous ply if it was searched at depth 21 or more.
Stockfish PA GTB has forward propagation feature too. That means, if one set, for example, depth 20, then all PV positions with depth 20 or more are saved. If engine reaches depth 25 it saves 6 positions. It's interesting to note the fact Baron saved all position and scores (after propagation) at game end, Critter saves them after a move is made, Stockfish PA GTB saves all suitable entries while on search. It means much more positions are saved.
The most effective way to use Stockfish PA GTB could be, apart from manually making moves and taking them back, to perform backward game analyses with fixed depth parameter. Fixed depth because we are sure score is propagated (with a Min-Max process) from game end to an opening stage.
It can probably still improved by introducing a score limit option, to avoid heavily won/lost position to be saved, or to prune the hash file better. A pity this project is dead.
F.S.I. Chess Teacher
, but I am glad that you are back to computer chess, because I found this thread really a fascinating one.