Thanks Marco!mcostalba wrote:Firstly I would like to congratulate with you and your employer for this fantastic achievement. It is really an important day for neural net technology.matthewlai wrote: It is now available on the DeepMind page. No pirating required!
It would be interesting to see how NN fares against chess champions.
Of course I don't mean humans, but traditional chess engines.
Indeed I think that chess is basically a computer oriented game, not a human oriented game.
In the 60' it was picked up as an AI test, perhaps because was seen as a low hanging fruit that 'resembled' what humans were supposed to excel. Of course today this seems a very naive approach considering that humans compared to computers excel in totally different topics than playing chess, but at the time other "human kind" topics as natural language, vision, etc.. were probably too far away from the possibilities of the time and so the chess 'trick' seemed a nice marketing vehicle to attract some attention (and funds) in this new AI research thing.
Today we know that traditional chess engines, i.e. engines programmed with a well known bag of typical computer science algorithms can easily outperform the world champion, and even on a mobile (as long as the batteries last
What we don't know is if an engine that leverages NN for evaluation and search heuristics can outperform the careful hand-engineered and well tuned features and terms of a standard chess engine.
I hope that a NN chess engine will beat Stockfish one day, this would be the little revenge that humans take against computers!
I didn't actually work on AlphaGo (I only joined recently), but it's a very exciting time to join, and I am honoured to be working with so many incredibly talented people.
I believe ML will eventually take over in chess as well, just not sure when. Chess engines still have much higher branching factors than humans. I believe that is due to the inflexibility of transposition tables, for the most part.
Humans don't need exact matches of positions to reuse results from previous searches (or branches of searches), and that gives humans a huge advantage.
In the computer chess world, that would be equivalent to a sub-position-level transposition table, that can remember useful concepts or patterns from previous searches. I don't see how that can really be implemented reliably in a traditional engine - it's way too abstract.
Machine learning, however, may have an answer. It's difficult, but I think some of the more recently published exotic NN architectures, as well as unpublished stuff, can have a decent shot. It's just a question of whether anyone will devote the time and effort to do this, when there are millions of very interesting machine learning problems out there.