Cerebellum analysis of the AlphaZero - Stockfish Games

[Thomas Zipproth]

I calculated the AlphaZero games and added them to Cerebellum, which took some time because Stockfish naturally needed very long calculation times and many added nodes to finally get the (most likely) correct results.

http://www.zipproth.de/Brainfish/Cerebe ... aZero.html

I thought about a way to present a game calculation in a most comprehensive but simple way, and finally ended up with this images.
Of course all side lines are missing, only the best move is additionally displayed, if it is different from the move which was played in the game.

Some conclusions or assumptions gained from the results:

Stockfish's evaluation ist correct most times, but it does not distinguish between a position which is dead draw and other positions which are theoretically also draw, but very complicated with many occassions to make a wrong move. In such positions stockfish sometimes makes errors leading to a loss, even if most times the evaluation of close to zero is correct. AlphaZero has a higher positional understanding and doesn't need scores. It always can play the move with the best winning chances, even if the position is theoretically draw, a case in which Stockfish only sees zero scores for all possible moves.

The AlphaZero neuronal network works like a book (for example Cerebellum) in the opening, because it (most likely) did not learn only positional factors but openings too. Stockfish has to calculate the opening completely new for every game, which is nearly impossible at normal thinking times.

AlphaZero doesn't seem to make big tactical errors which is astonishing in first sight, but can be explained by the search algorithm it uses. Combined with the neuronal network move selection it doesn't seem not have the problems of the Stockfish pruning because the neuronal network makes very few pruning errors.

Some additional results from the Cerebellum calculations:

In longer and deeper calculations the score often becomes simple zero, which make it impossible to select the move with the highest winning chances.

The tempo factor of about 0.08 sometimes seems to interfere with the evaluation, which means that in many cases longer lines are ending up in positions with absolut no positional advantage but a score of 0.08, which somehow seems to result from the tempo advantage.

[corres]

[quote="Thomas Zipproth"]
Stockfish's evaluation ist correct most times, but it does not distinguish between a position which is dead draw and other positions which are theoretically also draw, but very complicated with many occassions to make a wrong move. In such positions stockfish sometimes makes errors leading to a loss, even if most times the evaluation of close to zero is correct. AlphaZero has a higher positional understanding and doesn't need scores. It always can play the move with the best winning chances, even if the position is theoretically draw, a case in which Stockfish only sees zero scores for all possible moves.
[/quote]

This phenomenon is a consequence of the working of AlphaZero.
It uses to evaluate positions vectors and probabilities. These are depend on the relative positions of the pieces and the winning chance of their moving. Their "value" of the positions contain more information than the value of positions in centipawn used by normal engines like Stockfish.
But I think position evaluation of Stockfish may overhaul using milipawn and better optimization.

[quote="Thomas Zipproth"]
The AlphaZero neuronal network works like a book (for example Cerebellum) in the opening, because it (most likely) did not learn only positional factors but openings too. Stockfish has to calculate the opening completely new for every game, which is nearly impossible at normal thinking times.
[/quote]

Really, AlphaZero basically a kind of "book maker" with enormous power.

[quote="Thomas Zipproth"]
AlphaZero doesn't seem to make big tactical errors which is astonishing in first sight, but can be explained by the search algorithm it uses. Combined with the neuronal network move selection it doesn't seem not have the problems of the Stockfish pruning because the neuronal network makes very few pruning errors.
[/quote]

Against Stockfish they used not only the NN with an appropriate interface but a chess engine too. But what a chess engine - this is the main question. According to experts it is dubious to use an engine based on MCTS search.

[quote="Thomas Zipproth"]
Some additional results from the Cerebellum calculations:
In longer and deeper calculations the score often becomes simple zero, which make it impossible to select the move with the highest winning chances.
The tempo factor of about 0.08 sometimes seems to interfere with the evaluation, which means that in many cases longer lines are ending up in positions with absolut no positional advantage but a score of 0.08, which somehow seems to result from the tempo advantage.
[/quote]

This is the obvious disadvantage of the positive contempt.
Like strong enemy one ought to set the contempt to zero or minus value.

[Leo]

Could Brainfish with a massively trained Cerebellum compete with Monster Zero I mean Alpha Zero?

[shrapnel]

People still thinking AlphaZero won because of some of kind of massive Book !
Really sad, to see how even intelligent people blinker themselves to the Truth, when it doesn't it doesn't suit them.
All that DeepMind wrote in their Paper about Reinforcement Learning and MCTS Algorithm didn't make the slightest dent in their way of thinking.
No point arguing with such rigid, inflexible minds.
Just sad.

[Milos]

People still thinking AlphaZero won because of some of kind of massive Book !
Really sad, to see how even intelligent people blinker themselves to the Truth, when it doesn't it doesn't suit them.
All that DeepMind wrote in their Paper about Reinforcement Learning and MCTS Algorithm didn't make the slightest dent in their way of thinking.
No point arguing with such rigid, inflexible minds.
Just sad.

You know what is sad, that you go around constantly trolling and insulting ppl and mods do absolutely nothing about it...

[cc2150dx]

People still thinking AlphaZero won because of some of kind of massive Book !
Really sad, to see how even intelligent people blinker themselves to the Truth, when it doesn't it doesn't suit them.
All that DeepMind wrote in their Paper about Reinforcement Learning and MCTS Algorithm didn't make the slightest dent in their way of thinking.
No point arguing with such rigid, inflexible minds.
Just sad.

You know what is sad, that you go around constantly trolling and insulting ppl and mods do absolutely nothing about it...

+1. The way I look at it, If you have nothing good to say, don't say anything at all.

[corres]

[quote="corres"]

[quote="Thomas Zipproth"]
Some additional results from the Cerebellum calculations:
In longer and deeper calculations the score often becomes simple zero, which make it impossible to select the move with the highest winning chances.
The tempo factor of about 0.08 sometimes seems to interfere with the evaluation, which means that in many cases longer lines are ending up in positions with absolut no positional advantage but a score of 0.08, which somehow seems to result from the tempo advantage.
[/quote]

This is the obvious disadvantage of the positive contempt.
Like strong enemy one ought to set the contempt to zero or minus value.[/quote]

During analysis of games is also disturbing that Tempo behaves as Contempt. So developers of Stockfish should get the min(Tempo, Contempt) as Tempo. It works well.

[Lyudmil Tsvetkov]

I calculated the AlphaZero games and added them to Cerebellum, which took some time because Stockfish naturally needed very long calculation times and many added nodes to finally get the (most likely) correct results.

http://www.zipproth.de/Brainfish/Cerebe ... aZero.html

I thought about a way to present a game calculation in a most comprehensive but simple way, and finally ended up with this images.
Of course all side lines are missing, only the best move is additionally displayed, if it is different from the move which was played in the game.

Some conclusions or assumptions gained from the results:

Stockfish's evaluation ist correct most times, but it does not distinguish between a position which is dead draw and other positions which are theoretically also draw, but very complicated with many occassions to make a wrong move. In such positions stockfish sometimes makes errors leading to a loss, even if most times the evaluation of close to zero is correct. AlphaZero has a higher positional understanding and doesn't need scores. It always can play the move with the best winning chances, even if the position is theoretically draw, a case in which Stockfish only sees zero scores for all possible moves.

The AlphaZero neuronal network works like a book (for example Cerebellum) in the opening, because it (most likely) did not learn only positional factors but openings too. Stockfish has to calculate the opening completely new for every game, which is nearly impossible at normal thinking times.

AlphaZero doesn't seem to make big tactical errors which is astonishing in first sight, but can be explained by the search algorithm it uses. Combined with the neuronal network move selection it doesn't seem not have the problems of the Stockfish pruning because the neuronal network makes very few pruning errors.

Some additional results from the Cerebellum calculations:

In longer and deeper calculations the score often becomes simple zero, which make it impossible to select the move with the highest winning chances.

The tempo factor of about 0.08 sometimes seems to interfere with the evaluation, which means that in many cases longer lines are ending up in positions with absolut no positional advantage but a score of 0.08, which somehow seems to result from the tempo advantage.

In fact, your observation Alpha used an opening book is correct, why should not they simply acknowledge it?
But then, what is left from their achievement?

We don't know if Alpha makes no big mistakes, it makes many, just that most of us are too weak to catch those, while weaker SF will provide no assistance.
I am certain the games are full of Alpha mistakes, just no time to analyse closer now.

If SF and Komodo make suboptimal moves/mistakes at least every second move and Alpha is just tiny but stronger, then its quality of play is comparable and it makes and awful lot of mistakes.

I even don't want to think how many mistakes Alpha makes...

[Lyudmil Tsvetkov]

People still thinking AlphaZero won because of some of kind of massive Book !
Really sad, to see how even intelligent people blinker themselves to the Truth, when it doesn't it doesn't suit them.
All that DeepMind wrote in their Paper about Reinforcement Learning and MCTS Algorithm didn't make the slightest dent in their way of thinking.
No point arguing with such rigid, inflexible minds.
Just sad.

But they are checking facts, while you are just speculating.

[Ras]

But they are checking facts, while you are just speculating.

Btw., these facts somewhat clash with your claim that 80% of the games were lost right in the opening.

Not qualifying: games 1, 2, 3, 7, 8, 10.
Qualifying: game 4.
Debatable: games 5, 6, 9 (move 18/19 is not really opening anymore, IMO).

Or, that is the alternative conclusion, the opening book wouldn't have changed much because it doesn't see the early losses that you can see.