Google Deepminds Alpha Zero beats Stockfish

AlphaZero during its learning process may very well have encoded in its neural network information which amounts to having an "opening book" and an "endgame tablebase". Stripping Stockfish of these then certainly amounts to an unfair advantage.

Having said that, I have always believed that claiming that chess is already "practically" solved ( to a draw ) since the best engines today can always keep a draw when starting from an equal position ( especially from the starting position ), is complacency and utter nonsense.

The fact that already with 7 pieces ( as the Lomonosov table base proves ) there are forced mates that take more than 500 moves ( even if there is a technical draw somewhere in them due to the 50 move rule ), is proof that chess can generate unbelievable amount of complexity.

Stephen Wolfram is considered by many a crank, but I cannot get rid of the thought that the basic message laid down in his book "New kind of science" is correct. This message is about computational irreducibility.

Any computationally universal system ( that is capable of computing any function that is computable in the sense of the Turing-Church thesis ) is also computationally irreducible. This means that you cannot make predictions about the evolution of the system without letting it play through the steps of its evolution. The only way you can know what a system will do, is to run it and see what it does. You cannot make any short computation that will let you predict this.

While chess is not a computationally universal system ( most obviously it is simply finite ) it behaves in many ways as if it was one. In particular it shows signs of computational irreducibility ( as the said long mating sequences show ).

To claim that you can tackle such a system with tree search combined with heuristics to a near perfect level is a big complacency. This goes against the very essence of computational irreducibility.

There indeed may be different approaches, such as a neural network, that may upset a conventional search engine, just because the fountain is very deep and conventional search just scratches the surface.

nice post @jatekos

I would like to actually see the rest of the games especially the draws i still cant beat Chessmaster any edition

@jatekos Looking at the TCEC games between the strongest engines at the moment and their yearly progress it is pretty clear that chess is far from being weakly solved now. (That is, it is still impossible to code an always-drawing player with a state-of-the-art hardware and algorithms). Corresponding chess players also know some limitations of chess engines and how to improve over them in some positions.

However, I'm not sure about its computational irreducibility. Forced mates in 500 are quite artificial positions. I don't think it is possible to create a strong solution of chess, that is, an optimal player from any position (equivalent of 32-men bases). But speaking about weak solution, the reasonable positions reachable from the starting may turn out to be pretty restricted and could be analyzed with a complicated euristics, like it was done for Antichess.

Urgent News:
AlphaZero learned how to make omlet without breaking an egg in just 3 hours working on 4TPUs. We can put it's win over Stockfish already aside, just because of that :)

Computational irreducibility is interesting, and it's known that generalized chess is EXPTIME complete. Solvability is interesting too. However, chess with the 50 move rule is also O(1) with a lookup table, but said lookup table can't be stored in our universe. These complexity analyses are great, but we still want to solve many problems in the face of staggering complexity with limited resources. That makes chess a great candidate for this.

This is what makes me excited about AlphaZero. It evaluates an order of magnitude less positions than Stockfish and still wins. AlphaZero does this by learning what patterns to extract from chess games and the heuristics, probably in the form of higher level patterns, and applies these heuristics without humans intervening.

This is also another victory of modern machine learning over old school AI. I find this to be very encouraging, despite the fact that this stuff is not yet fully understood.

It does use the TPUs, but this hardware is probably coming to your phone/tablet/laptop/PC in the next five years or even the next year. NVIDIA already has hardware that provides this stuff and people have been doing machine learning for the past five years at least with GPUs.