Comment by thomasahle | Hacker News Reader

Comment by thomasahle | Hacker News Reader

As the author of sunfish, I'd like to point out something about learning in chess, which is a topic that interests me a great deal.

When sunfish (with its just 111 python lines of chess logic) 'evaluates' a position, it uses the perhaps simplest known effective method: A piece-square table. The method ignores any interplay between pieces on the board and calculates the sum of Table[coord][type] for each piece on the table. E.g. a white knight on a1 may be worth 73 'units', and on f3 it may be worth 98 'units'. That's all there is to it. Any program which has greater than this level of precision, and equally precise searching, should be able to beat sunfish.

The above may sound naive - and it is - but actually most of the advanced ideas used in chess evaluation functions, can be generalized from this method. "Rook connection" is just a measure that includes two pieces instead of one, and "pawn shield" is the generalization to three pieces. Experiments with grandmasters reveal they "recall" positions in 'chunks' of connected pieces. And this memory is what they use to guide their search. (Papers like 'Perception in chess' and lots of newer research).

So, the role of machine learning in modern engines is to tune the parameters for evaluation and search pruning (deciding what positions are worth examining deeper). For the actual decision of which piece to move to where, you still need search algorithms to crunch millions of positions per second.

Sunfish is really impressive work. From my (brief) understanding of Sunfish, the evaluation function is essentially equivalent to a hardcoded 1 layer network in Deep Pink.

You're right that everything else equal, a better evaluation function should lead to a better chess engine. However in practice I think better evaluation functions means slower evaluation function. So there's some really interesting trade-off there. I doubt humans evaluate more than a few thousand positions, so it seems like a slow but more accurate evaluation function could play chess pretty well

One interesting line of research, I think, is using 1 or 2 layered networks to 'simulate' more complex evaluation functions. If you could train such a network to get within a 10% error of Stockfish's evaluation, then you might be able to distil that network as a faster evaluator to plug back into Stockfish for an even stronger engine. As you say, one hard problem is probably finding actually interesting positions to sample for the training.

Anyhow, it's fun to see how engines like these battle it out. It may also be that your approach can yield a more 'fun to play' engine for us mortals.

I think that's pretty useful approach. It's kind of similar to Hinton's latest work on model compression: http://www.ttic.edu/dl/dark14.pdf

The problem with deep models is when you end up having more than 1 hidden layers, you have a big matrix multiplication to get between the layers. If your hidden layers are a few thousand units, that's still pretty slow. Doing things in minibatches or on the GPU speeds it up significantly, but I'm guessing it's still orders of magnitudes slower than whatever Stockfish uses

Sure, the second layer would have to be very sparse. That makes sense since most multi-piece 'chunks' are not really that interesting.

Thanks for sunfish, great work!

Reply on news.ycombinator.com