My high-level takeaway from Matthew Lai's Giraffe Chess Paper is that one would want to use broad, shallow game trees, with some method of evaluating the probability of a favorable outcome for a given board position. Is this correct?

(Still working my way though the AlphaGo paper, but the method seems to be similar.)


If you mean high level assessment of self-learned evaluation functions in chess, then no, the advantage of a better evaluation function lies in the ability to prune the search tree more aggressively. So you would on the contrary try to search narrowly but deeply.

(In reality neural network based evaluation functions are so slow, that you would search narrowly and still not get very deep. Nor very strong.)

If you mean chess programming in general, than the answer is also no. In chess you have to go deep, at least selectively, because tactical possibilities that occur deep in some variations are important.

  • $\begingroup$ My thinking was that if you use narrow, shallow trees, you can potentially have more of them, per your note that even the deep trees are not going to be deep. (I'm looking at a model where I might ideally have 25 plies, but the numbers get pretty daunting, even if I'm pruning to 3 choices per ply.) $\endgroup$ – DukeZhou Nov 4 '16 at 14:44
  • $\begingroup$ What do you mean "more of them"? You always have one tree, rooted in the position you are starting from. $\endgroup$ – BlindKungFuMaster Nov 4 '16 at 14:49
  • $\begingroup$ Technically, that's true, but due to the unique nature of the game I'm working on, I'm finding it instructive to look different choices from a given game position as separate trees with different depth and pruning requirements. (I should mention that we're not working on Chess, but an entirely new non-chance, perfect information game method with completely different rules and requirements.) $\endgroup$ – DukeZhou Nov 9 '16 at 19:19
  • $\begingroup$ PS In my original comment I should have been more careful in my wording. What I meant was more branches that could be evaluated for some advantage at a strategic level, before drilling down to evaluate the local, tactical benefit or a given strategic choice. I do greatly appreciate your taking the time to answer and provide clarification! $\endgroup$ – DukeZhou Nov 9 '16 at 19:28
  • $\begingroup$ If you are working on a different game, then of course all bets are off. If the branching factor is very high and the positional assessment is very tricky, neural networks start to make a lot more sense. How exactly you search also depends on how good moves are distributed and how deep you have to look to recognise them. With a completely new game, you'll probably just have to try a lot of strategies. $\endgroup$ – BlindKungFuMaster Nov 10 '16 at 16:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.