The thing is, DM and normal engines evaluation is different. Deep mind doesn't need to bruteforce positions as much to achieve an advantage, hence why we have 70,000,000 positions for SF versus 80,000 from DM, and won't really matter hardware or hashsize since DM's genetic algorithm is only layered on learning phase, not when playing. And should be considered that DM had only four hours of training under 8 TPU's, it could've trained for days.
The four hours of training were deemed enough to go against all the database SF8 have been collecting since its conception. There is no comparison between them really. SF8 can be improved as much as they can, its going to be a losing battle.
"won't really matter hardware" lol :D good one, I guess AlphaCerebro can run then on my Pentium 133Mhz and destroy Stockfish running on some super computer, sounds fine, no :D since it learned and solved (have read it in telegraph article) chess :D
When one makes a claim like "Furthermore, the same algorithm was applied without modification to the more challenging game of shogi, again outperforming the state of the art within a few hours.", the hardware really matters. The ELO difference between SF and AlphaZero is of order 100, and this is a very sensitive order when you start to tweak opening knowledge, endgame knowledge, used version and especially hardware. (Doubling the calculation speed is 70-100 ELO increase in modern engines.)
Note also that "just 4 hours of learning" also needs a hardware context as the resources of Deep Mind are much, much bigger than fishtest framework powered by a bunch of volunteers will ever have. Also note that the skill graphs become flat quickly, thus the FINAL strength actually depends on the structure of their algorithm, not on these hours (the logic "If they run for over 9000 hours, Stockfish would be completely destroyed" just doesn't work). They are only the indicator of game complexity and practical effectivity of their algorithm.
They also repeat 3 times that Stockfish 8 is a TCEC-2016 champion which will make some people believe that it is really state-of-the-art now, but indeed not. There was more than a year of development since then, and a year of development in Stockfish team is always a big leap in ELO, also comparable with the differences they claim. Let alone ASMFish and Brainfish that were stronger even during TCEC-2016 but not actually participating due to special reasons.
Well, I believe that AlphaZero is a serious achievement in AI field, as a proof of concept that AI can teach itself to a solid level with almost zero prerequisite knowledge. However, its "outperforming of Stockfish" is a quite exaggerated news and its incluence on the chess engine world probably won't be serious. But who knows, though.
@no_bullet_thanks Google = Skynet
1=1
i win
How about Google develop a neural network that can detect engine users :D
this machine is dangerous i can believe how this machine play
I absolutely agree with
@Wolfram_EP [#5 and #35]. AlphaZero does indeed have the potential to change the future of AI engineering and maybe even chess engines/chess. There's a good chance that AlphaZero is stronger or even much stronger than Stockfish, but judging that by this match would be an extremely unfair assessment. The extents of the handicap are correctly explained by
@Wolfram_EP. It would be very interesting if a match b/w SF and AlphaZero was a set up with equitable conditions such as optimal hardware, opening books/tablebases for SF etc.
Tord Romstad, who is one of the creators of SF commented on the match and also answered some questions in a reddit thread which I've linked below.
www.reddit.com/r/chess/comments/7igro1/alphazero_reactions_from_top_gms_stockfish_author/dqyndi0/This is very interesting. I would love to see AlphaZero play Magnus Carlsen. Also, it would be nice to see AlphaZero play the strongest incarnation of StockFish.
Would I find interesting about this is that AlphaZero is trying to learn heuristics like humans do in order to perform better without having to make more calculations.
AlphaZero performed 80 thousand calculations per second compared to StockFish using 70 million calculations per second and still won!
This evidence makes me suspect AlphaZero still outperforms modern versions of StockFish.
If humans had the ability to compute like machines, I think we would beat engines like StockFish, because StockFish has innate biases programmed by humans and only rules based evaluations. An engine like this can only improve by humans finding other rules and optimizations to add.
An engine like AlphaZero can teach itself which rules to add without humans interfering. This is the power of deep learning, the ability to extract important features from the data and develop optimal strategies.
I do wonder how large the weights file is for this though. It might not be able to work inside of a phone. Updates to end-users might be difficult, since training AlphaZero with a new or modified architecture, or more epochs probably means you have to download the whole thing again.
Evaluation is probably quite fast, since it is the training that takes most of the time. On the downside, it might be necessary to have multiple GPUs in order to even run AlphaZero.
This also begs the question: let's see a modern Stockfish and AlphaZero go at it on 4 GPUs each and see which one wins. I bet AlphaZero wins, but it's only fair to give the same computational resources to each engine IMO.
Why are you guys questioning the fairness or conditions of the match? That is not important. It was just a little demonstration of Alpha Zero's potential. We're talking about a totally different technology, much more advanced. Go and learn a bit about Deep Learning.
There is no interest in AlphaZero playing Magnus Carlsen, a human will not manage to draw even a single game, I'm afraid. It would be nice to see a corresponding chess game against an ICCF world champion using Stockfish but this is unlikely to happen.
The number of nodes per second is quite a meaningless figure by itself. This is the number of nodes that appeared in a high-level search algorithm, but who knows what was going inside the neural network which evaluated them and which nodes it in a sense "looked at". Stockfish search is consisting of high-level search and quiescense search, and only the quiescense search can actually give a more or less reliable value of the position, but not a single eval. But neural network is probably figuring out the reliable value directly, sacrificing the speed. So probably it is more fair to compare the Stockfish nodes per second that are in the high level search rather than in all the search.
Evidence makes me suspect AlphaZero is not going to outplay Stockfish in the best conditions for the latter, but the validity of this fact is of computer chess sportive interest only and doesn't have any value for practical chess play, let alone for AI questions Deep Mind is interested in.
