AlphaZero Chess is not that strong ...

[mjlef]

I do not quite understand your graph, but is your program using LazySMP? Because hash size is going to have a large effect using that search scheme.

It seems to be time to fixed depth searches. But with LazySMP, time to depth suffers while the goes up.

[hgm]

The graph was for single thread. Indeed time to depth, and load factor defined as reported node count divided by number of table entries.

Whether SMP would behave significantly different is certainly worth investigating, but not so obvious to me that it can be assumed without evidence. With SMP, threads tend to work in the same sub-tree, which can easily be just 10% of the total tree, and thus still fit in the hash table even under conditions of large overloading.

[Werewolf]

Because hash size is going to have a large effect using that search scheme.

It seems to be time to fixed depth searches. But with LazySMP, time to depth suffers while the goes up.

It would be really nice to get to the truth on this. HG Muller claimed that going from 1GB to 32 GB only adds 7.5 elo which is quite a bit below what I had guessed (25 elo).

[mclane]

So a good book makes it impossible to see the difference between a 3500 and a 6000-Elo engine: all games will be draw. To test engines it is thus essential to play them without a good book. Tests with a book are meaningless; they say nothing about the engine, but only something about the book.

What is the sense of creating a big searchtree like stockfish / Komodo / Houdini are doing so, if the outcome is that the engine has almost no clue what is going on on the board although it was computing 23-28 searches deep in the middle game.

Stockfish has material on board that gets 3,5 and more pawns value but the pieces cannot move !!

Pieces that cannot move are worthless.

Stockfish ate almost if not all sacrifices by AZ.

This is the indicator that it had no clue what was going on.

Although „evaluating“ 70.000.000 nodes per second.

Either the evaluation is completely wrong then or the search tree is build in a completely wrong and wasted area or the second is a result of the first.


What we need now, is IMO a complete NEW cut and new ideas.

We cannot continue like this.

We thought this is the way to go in the last 50 years.
But it seems this is a dead end street.

[Werewolf]

We thought this is the way to go in the last 50 years.
But it seems this is a dead end street.

Of course it's not a dead-end. Engines are improving at 30 elo a year - that is PROGRESS.

Just because they can't play perfectly yet isn't the same as saying there's no progress.

[hgm]

It would be really nice to get to the truth on this. HG Muller claimed that going from 1GB to 32 GB only adds 7.5 elo which is quite a bit below what I had guessed (25 elo).

Well, take Stockfish, take a machine with quie a few cores (8 threads would already give a good indication), make a graph like the one I posted, and you would know.

[mclane]

Did you take a look into the game and in stockfishs Evaluation during its games ?!

[FWCC]

Alphazero IS THAT STRONG I think it is revolutionary

[fern]

I do not see the point of trying to minmize this new monster feat compared with the stuff we already know. The essence of the matter if that this new criature is capable of learning chess and playing at SGM level after a couple of hours of learning from zero.
If it is not better than Stockfish or any other no doubt it will be in the next iteration.
We are in presence of something new.

Fern

[mario carbonell]

I do not see the point of trying to minmize this new monster feat compared with the stuff we already know. The essence of the matter if that this new criature is capable of learning chess and playing at SGM level after a couple of hours of learning from zero.
If it is not better than Stockfish or any other no doubt it will be in the next iteration.
We are in presence of something new.

Fern

That is the point.

AlphaZero is the first iteration of a new technology.

Somebody will continue their work and do something even stronger. Maybe in a few years every chess engine will be derived from AlphaZero.

Another interesting thing that can be done with neural networks is style transfer. Maybe it could be trained to imitate the style of you favorite chess player, could be Mijail Tal, Anatoly Karpov, Gary Kasparov etc...