[sytelus]

It's surprising that it is able to win 1/3rd of the time. The problem here is that input does not lie in any continuous space. I mean, you may have 1 billion board states in your training but is it possible to interpolate values of other states using this? For example, for one vector representing certain board state, even a slight change may have completely different outcome. I would think most learning methods, including deep learning, would excel when there is some sort of interpolatable continuity in inputs on which prediction is desired. Therefore the challenge would be transform discontinuity in one board state to another to more continuous space.

[phreeza]

Doesn't this hold true for other domains where deep learning seems to be successful, for example natural language processing? My intuition would be that initial layers of a deep network actually learn smooth-ish representations of the input space.

[V-2]

My thoughts exactly. As an exercise it's surely interesting, but neural network approach is inherently unsuitable for chess.

Chess requires 100% accuracy and in which just because positions are similar, doesn't mean that best moves in these positions have to be in any way similar too.

On the other hand, it sort of mimics the way human player thinks, in terms of recognizing certain patterns. After all, even grandmasters do not bruteforce their way through all possible combinations. We use a hybrid approach: recognize certain strategic patterns first (to drastically narrow down the search tree), and perform calculations on the top of that.

Chess engines can wipe the floor with any player where tactics is involved; the trick of beating a computer is to close the game and take advantage of the fact that it's not able to formulate a long-term PLAN (whose consequences are beyond its horizon).

See how Nakamura repeatedly beat Rybka in blitz games a few years ago, eg.: http://www.chessgames.com/perl/chessgame?gid=1497429 - very instructive :)

[sushirain]

> just because positions are similar, doesn't mean that best moves in these positions have to be in any way similar too.

To alleviate this, one can add more abstract/heuristic information about the position to the input (indicators for complex relations between several pieces). This kind of high-dimensional vector would be more robust to small changes, and make the objective function more smooth. Perhaps the non-linearities introduced by the three layers cannot do this as effectively.