AlphaZero - Tactactical Abilities

[pinkfloydhomer]

Sorry, I am too lazy to search myself. Was anything more conclusive ever published regarding my initial question?

There is a lot of noise in this thread. The most concrete counter argument is from H.G. Muller: that the tree searched by az is much smaller.

To that I would say this:

1) Yes, that does represent some kind of breakthrough, but

2) You could always use an immense hardware advantage to do much more selective searching/pruning using more expensive heuristics, NN-based or not, searching much fewer nodes but much deeper in the right ones, the tactically decisive ones. In that sense, stockfish was then still, as I proposed, out-searched, out-tactic'ed, moreso than az having some brilliant positional understanding. As I wrote, even a slight advantage in tactical ability will look like a better positional understanding.

And finally: An algorithmic achievement, such as a better tactical or positional understanding coming from a tactical search coupled with an NN assessment, cannot be deemed to be a breakthrough or at least to be "better" without looking at either equal hardware or big O complexity analysis or similar. If you create an algorithm that is "better" than some other algorithm but it requires a million times the hardware power or money, is it really a better algorithm? Or just better hardware?

NN and deep learning and modern AI techniques in general are powerful, no doubt, especially since they can be applied fairly easily to a lot of different domains, and these techniques will help humans to solve a lot of problems better in the future. It will probably also eventually make us able to create even better chess engines given the same or even modest hardware at some point.

But the achievement of az seems to me to mostly be a triumph of hardware.

[Terry McCracken]

You do know you're comparing apples to oranges don't you?
Alpha Zero is not a chess engine, it's not a chess program.
It can play countless games and it teaches itself from scratch. It is only given the rules and how the game is played. In chess, the moves.
It can beat any program or many players at once. You can't simply out search it.
It has a form of intelligence that other computers lack. If you want to out smart Alpha Zero then you will have to build a better faster Neuro Network and learning algorithm. I think this type of AI is a turning point in computers.

[pinkfloydhomer]

Did you read the first post and the rest of the thread?

[wickedpotus]

You do know you're comparing apples to oranges don't you?
Alpha Zero is not a chess engine, it's not a chess program.
It can play countless games and it teaches itself from scratch. It is only given the rules and how the game is played. In chess, the moves.
It can beat any program or many players at once. You can't simply out search it.
It has a form of intelligence that other computers lack. If you want to out smart Alpha Zero then you will have to build a better faster Neuro Network and learning algorithm. I think this type of AI is a turning point in computers.

I think the MCTS search in A0 is no match for the natural stupidity of some humans (Natual Intelligense).... grain of salt..

[Ovyron]

You do know you're comparing apples to oranges don't you?
Alpha Zero is not a chess engine, it's not a chess program.
It can play countless games and it teaches itself from scratch. It is only given the rules and how the game is played. In chess, the moves.
It can beat any program or many players at once. You can't simply out search it.
It has a form of intelligence that other computers lack. If you want to out smart Alpha Zero then you will have to build a better faster Neuro Network and learning algorithm. I think this type of AI is a turning point in computers.

This comment aged really poorly, Alpha Zero could easily be beaten by Stockfish NNUE on today's common man's PC.

[dkappe]

You do know you're comparing apples to oranges don't you?
Alpha Zero is not a chess engine, it's not a chess program.
It can play countless games and it teaches itself from scratch. It is only given the rules and how the game is played. In chess, the moves.
It can beat any program or many players at once. You can't simply out search it.
It has a form of intelligence that other computers lack. If you want to out smart Alpha Zero then you will have to build a better faster Neuro Network and learning algorithm. I think this type of AI is a turning point in computers.

This comment aged really poorly, Alpha Zero could easily be beaten by Stockfish NNUE on today's common man's PC.

Except that Stockfish NNUE is a “better, faster Neural(sp?) Network.” The learning algorithm is ballpark the same speed. And it did mark a turning point in computer chess: all top engines are using neural networks of one form or another.

[Ovyron]

Except that Stockfish NNUE is a “better, faster Neural(sp?) Network.” The learning algorithm is ballpark the same speed. And it did mark a turning point in computer chess: all top engines are using neural networks of one form or another.

Yeah, but NNUE technology is completely different from Leela NN technology, who knows if Hisayori Noda could have come with it independently. NNs had been used for all sort of things in the past, it was a matter of time it was chess's turn, though I had expected it'd kill chess, but we'd have reached this point in 2 years with classical eval anyway (we're just having the level of what would have been 2024's Stockfish.)

And the zero approach didn't even work,the strongest NNUE nets were done with reinforced learning, and then with the most deep and accurate evals, nothing to do with learning chess from scratch at all.

At this point I feel that all the time and effort spent on Leela was wasted time, and the best thing that came with it was the slight improvement that it gave to Stockfish when it used its data. Otherwise, only Nodchip can tell us.

[dkappe]

Except that Stockfish NNUE is a “better, faster Neural(sp?) Network.” The learning algorithm is ballpark the same speed. And it did mark a turning point in computer chess: all top engines are using neural networks of one form or another.

Yeah, but NNUE technology is completely different from Leela NN technology, who knows if Hisayori Noda could have come with it independently. NNs had been used for all sort of things in the past, it was a matter of time it was chess's turn, though I had expected it'd kill chess, but we'd have reached this point in 2 years with classical eval anyway (we're just having the level of what would have been 2024's Stockfish.)

And the zero approach didn't even work,the strongest NNUE nets were done with reinforced learning, and then with the most deep and accurate evals, nothing to do with learning chess from scratch at all.

At this point I feel that all the time and effort spent on Leela was wasted time, and the best thing that came with it was the slight improvement that it gave to Stockfish when it used its data. Otherwise, only Nodchip can tell us.

1. NNUE is an implementation of a very simple, shallow neural network architecture that could be evaluated efficiently. As a neural network, it’s much more primitive and harkens from an earlier time than the convolutional neural networks used by Alpha Zero.

2. Reenforcement learning in the case of leela is from scratch. It starts with only the rules of the game and gradually learns to play good chess by playing against itself. Stockfish uses the leela data to train its nets. Whether that’s still “zero” is matter for debate.

[Ovyron]

1. NNUE is an implementation of a very simple, shallow neural network architecture that could be evaluated efficiently. As a neural network, it’s much more primitive

And much more effective. Actually this speaks of a much higher required intelligence to come with something where more can be done with less, all the people at Deep Mind couldn't do it.

2. Reenforcement learning in the case of leela is from scratch. It starts with only the rules of the game and gradually learns to play good chess by playing against itself. Stockfish uses the leela data to train its nets. Whether that’s still “zero” is matter for debate.

What is not for debate is that NNUE technology was much more successful than Leela's, and it was so much earlier than when Stockfish used Leela's data.

[smatovic]

1. NNUE is an implementation of a very simple, shallow neural network architecture that could be evaluated efficiently. As a neural network, it’s much more primitive

And much more effective. Actually this speaks of a much higher required intelligence to come with something where more can be done with less, all the people at Deep Mind couldn't do it.
[..]

https://www.chessprogramming.org/Giraffe

https://www.chessprogramming.org/Matthew_Lai

DeepMind never released a chess engine for the end user to run on a PC, that was simply not their goal, they cracked Go, Shogi and Chess with AlphaZero, mission accomplished, next target.

--
Srdja