Deep Learning Machine Teaches Itself Chess in 72 Hours, Plays at Master Level

technologyreview.com

3 pointsbetolink2 comments

I remember my AI class working with chess engines, they all had to evaluate game trees using one or another heuristic strategy. The problem is that the search space is huge and so far a machine can only beat a Grand Master by evaluating millions of positions in advance... i.e. Deep Blue. As an example that these(brute force) techniques do not scale was given by John McCarthy http://www-formal.stanford.edu/jmc/reti.html This is why this could be quiet a breakthrough on chess AI and possibly other domains.

From the paper, I gather that Giraffe uses standard alpha-beta search with a heuristic for position evaluation. The learning applies only to the heuristic. I.e. they did not train a learner to point to the next best move, they trained one to evaluate the position.

This is very impressive work considering it's a master's thesis (meaning it was done in a short period of time, not that master's students aren't capable of producing top-quality work), but: 1) applying learning/tuning to position evaluators is not new (the network is deeper than usual, though it also uses features that are known to be useful, and doesn't rediscover everything), 2) the rating gap when compared to state-of-the-art engines is absolutely massive (contrary to the article's claim that it is competitive). Giraffe is a great project and it is better than almost all humans, but it doesn't seem (yet!) to be a revolution in computer chess.

Reply on news.ycombinator.com