Alpha0 opening book is actually already stored in its network weights. It would most probably played weaker if you forced it to play from a real opening book.Branko Radovanovic wrote:Why do you think SF should have played with an opening book and an EGTB? AlphaZero had none. As I gather, it used actual computation to play all its moves, rather than relying on instant lookup.
That's a bad excuse. If Google didn't have time or know how to optimally implement time management, there is no reason to handicap other engine.SF surely lost 20-30 Elo at least due to fixed move time, but given the fact that AlphaZero itself played under the same conditions, there is no reason to believe the outcome would have been any different even with dynamic time management given to both sides. (Actually, if there is any mention of AlphaZero's time management in the paper, I've missed it.)
Moreover, not going into too much details, classical engine time management is more complex because of the nature of alpha-beta search, when failed high/low engine simply needs much more time to do the research of a position. This is not the case with MCTS.
That is not actually true. BTU that google is exclusively using is much more exotic than anything that exists. FPGAs are quite common and used.AlphaZero's hardware is better described as novel than specialized. There is no doubt it will become commonplace, perhaps in a matter of years, while FPGAs will remain exotic, however powerful and energy-efficient they might be.
What lacks is actually engineers who could use them efficiently.
Everyone today can learn to make some deep learning in Python in few hour following a YouTube course. Learning how to efficiently trranslate C++ to FPGA takes years. That's the main difference.
This is all fine, but MCTS is also quite an old concept in chess programming and used since forever in Go and other games (it had also fair share of usage in chess quite some time ago). And Alpha0 without it would be nowhere near to actual performance. So yes the way they organized the features and trained the network is quite novel and certainly a breakthrough, but none of that would have been possible without monstrous hardware and "old" techniques such as MCTS an reinforcement learning (at least 25 years old concept in chess).If I had to name just one reason why AlphaZero playing chess the way it plays is a monumental achievement, it is this: in the chapter title "Anatomy of a Computer Chess Program" the authors list a number of computer chess techniques (alpha-beta search, material imbalance tables, PSQT, mobility, pawn structure eval, king safety, QS, pruning, extensions, history, SEE, heuristics, TT, etc.) and conclude AlphaZero uses none of them. Not one. It appears that the only thing SF and AlphaZero have in common is the move generator. That's mind-boggling, and remains mind-boggling even if Google's results do appear to be oversold a bit.