CheckersGuy wrote:Yeah 100 elo above stockfish isn't like what they have accomplished with Go. However, I wonder how strong alphaZero could get given more training and optimizations specific to chess. Since AlphaZero was, basically, for any board game I think there might be a lot of improvements.
Let's bet the improvements will be 0.0 in 5 years' time.
This is their peak.
You can not achieve an engine much stronger than 2800 with such algorithms.
mario carbonell wrote:One possible explanation it that chess engines are near the maximum strength possible, that may be around 3600 or 3700 elo.
We can see that in the game of go, AlphaZero was still learning, but at chess seemed to have arrived to a plateau. Maybe the maximum strength?
That also speaks about the very good work done in Stockfish, that is very near to A0.
The problem with chess is that the draw result exists, and even one or two pawns difference in material is not enough to win.
Did not we think the Kasparov level, 2850, was about the maximum one could achieve?
Chess will NOT be solved before 5500 elos are attained, I guess even 7000.
My proofs?
Well, I am able to get better positions than SF almost everywhere, in open and closed games alike, just that I am not able to win all. SF makes suboptimal moves quite often, maybe every second or even every 2 out of 3.
mario carbonell wrote:One possible explanation it that chess engines are near the maximum strength possible, that may be around 3600 or 3700 elo.
I think they reached "the maximum strength possible" of this configuration of the neural network. But it can be improved !!
In what way?
If it learns by itself?
Did not they say they are NOT touching the code at all?
And then what will they change?
I thought the assumption was you give the engine the raw rules of chess, and nothing else, and it starts learning.
Well, that is what it managed to learn.
They could use a more powerful neural network, which is able to learn more. The NN they used now was not adapted very well to Chess. E.g. there were no patterns along board rays, while these are kown to be important for Chess due to the presence of sliders (pins, X-ray protection, discovered threats). These had all to be built by cascading local 3x3 patterns in consecutive layers, which takes a lot of learning. Using the same hardware, just configuring it better for Chess, they could greatly increase the capacity and learning speed of the NN. Doing 4 hours of training with a more powerful NN starting from scratch would be much more productive than putting more training hours in the current network.
Of course they could also switch to a hybrid approach, using conventional preprocessing of Chess positions to extract patterns that are considered important by Chess players (such as the mentioned pins, X-rays and discovered attacks, but also insufficient protection, overloaded protectors, attackers on the King neighborhood, various kind of Pawns). So that the NN could directly use those if it considers them useful, and only has to learn how much they are worth, or how they have to be modulated by other, yet to learn patterns, rather than having to discover (and later compute at great cost) these 'standard' patterns by itself.
There is so much they could try to do better. Or that we could do to try better...
Last edited by hgm on Tue Dec 19, 2017 8:44 pm, edited 1 time in total.
hgm wrote:They could use a more powerful neural network, that is able to learn more. The NN they used now was not adapted very well to Chess. E.g. there were no patterns along board rays, while these are kown to be important for Chess due to the presence of sliders (pins, X-ray protection, discovered threats). These had all to be built by cascading local 3x3 patterns in consecutive layers, which takes a lot of learning. Using the same hardware, just configuring it better for Chess, they could greatly icrease the capacity and learning speed of the NN. Doing 4 hours of training with a more powerful NN starting from scratch would be much more productive than putting more training hours in the current network.
But is not the NN what Alpha learned?
Why do you say 'to start with a new NN'?
mario carbonell wrote:One possible explanation it that chess engines are near the maximum strength possible, that may be around 3600 or 3700 elo.
That can be ruled out both for the quick self-play games during training and the 1 sec/move games used for obtaining the data that formed the basis for the graphs in Figure 1.
There is no doubt that AlphaZero at 1 min/move plays significantly stronger than AlphaZero at 1 sec/move. This is already visible from the graph, which shows it to be just a little bit stronger than SF when both play at 1 sec/move whereas it is considerably stronger than SF when both play at 1 min/move.
Ergo, AlphaZero's play at 1 sec/move has not reached some sort of "maximum strength possible".
More generally, such claims of the existence of a "maximum strength" not far from current top-engine strength should be treated with considerable caution. I am not aware of any experiment that showed a doubling of thinking time, however long, to gain only a very minimal amount of Elo.
AlphaZero is very strong. The best proof is this whole debate that in better conditions Stockfish could be on par with it. This is not exactly definition of weak. All points about SF having not optimal setup are true but they cannot change the fact that AZ is amazingly strong chess player.
All this is not any news for people following AlphaGo development. News about AlphaGo playing on professional level broke around 2 years ago and it only got stronger since then. If anything it was surprising that they didn't tackle chess earlier. They surely tried but it only shows how good current top engines are.
Seeing what happened with AlphaGo we should not expect much more from Google at least not until they have some kind of breakthrough. They will probably publish some more games, including self-played (maybe 20-50, it's not very likely to see more). Computer chess community should start their own project to see what's possible with NN approach. Merged with domain-specific knowledge it could end up maybe even few hundred Elo stronger than AZ, who knows. It's really exciting. In Go we can already have top human level program running on ordinary PC with modern GPU and very strong amateur level running on measly laptop.
Last point: AlphaZero approach works, denying it won't change anything. Look at Leela Zero project. With much smaller network and community effort, in one month it gained around 4000 Elo above random mover from self-played games only. Now it's around 10-12 kyu level I believe. It's not very strong yet but it's amazing anyway taking into account that it still plays weird opening moves and has some major tactical blind spots. When they are corrected (with more games and more training), it will play on dan level very soon.
Last edited by Dariusz Orzechowski on Tue Dec 19, 2017 9:00 pm, edited 2 times in total.
Lion wrote:What ever the conditions, just look at all the amazing moves/sacrifices A0 played that current other engines are not capabl off.
So one can complain of time, opening book or other but the games are there and speak for themselves!
Rgds
Well, all too often in the past, the TCEC champion has been crowned with +10 wins more, or close to that, out of 100.
Alpha has +30 or so, so just the difference between SF 6 and SF 8.
Why no one ever claimed SF 8 was an incomparable break-through?
Lyudmil Tsvetkov wrote:But is not the NN what Alpha learned?
Why do you say 'to start with a new NN'?
The NN is a machine (a virtual one, in this case, simulated by a TPU), or actually a class of (programmable) machines, like computers. Just like there are different types of computers (PC, Mac, tablets, phones) there are different designs of NN. A NN learns what to do by training it, similar to a computer does learn to do things by explicit programming. So the knowledge leared by AlphaZero is not the NN, but the program learned by and contained inthe NN (the billions of 'weights').
What I propose is to just upgrade the NN to a better one, like you would upgrade from a smartphone to a server PC, and than train it to contain much more knowledge than the current N of AlphaZero can hold or acquire in reasonable time.
