Yes, and if it had opening book or let to play by itself in the opening, it would have lost 0 as Black.
Objectively Speaking, Is Magnus a Patzer Compared to StockFish and AlphaZero?
Sort:
Dec 27, 2017
0
#263
Btw, Elroch has repeatedly stated that neither player had access to a database. This is false. AlphaZero had access to the games it had previously analyzed. At 80,000 moves per second and an average of 40 moves per game, it had 240,000 games stored in its database.
Dec 27, 2017
0
#264
You are incorrect. And AlphaZero played a few tens of millions self play games to learn its "evaluation function" (the neural network). More precisely, 700,000 minibatches of 4096 positions, i.e. about 2.9 billion positions.
(Thanks to CM ilmago for correcting a mental calculation error in the first version of this post!)
Dec 27, 2017
0
#265
In what way am I incorrect, in the size of the number of games AlphaZero played before the match? Ok, it was several million games, not 240,000. That makes it more lopsided, not less.
SmyslovFan hat geschrieben:
Btw, Elroch has repeatedly stated that neither player had access to a database. This is false. AlphaZero had access to the games it had previously analyzed. At 80,000 moves per second and an average of 40 moves per game, it had 240,000 games stored in its database.
And you think it becomes more true if it is repeat again and again? There is no database, and to store some games leads to nothing. And it is bullshit for older games which are played by weaker versions. You can store endgames with six or seven pieces, but this is not necessary for AZ (but note that AZ is not able to play the best line in each end game position, and it cannot give you the instant answer about a mate in 40 moves or something like that).
Dec 27, 2017
0
#267
SmyslovFan wrote:
In what way am I incorrect, in the size of the number of games AlphaZero played before the match? Ok, it was several million games, not 240,000. That makes it more lopsided, not less.
This should not be difficult for you. AlphaZero only has an evaluation routine which it learnt from self-play and a search method which uses that evaluation routine. The games it has played are not remembered, only the evaluation routine, defined by a large number of learnt parameters.
Dec 27, 2017
0
#269
SmyslovFan wrote:
Your condescension is noted.
It's not actually condescending to point out that someone should have no problem comprehending what AlphaZero does not do. I can see how anyone could be surprised that what it did do was so effective and that that might suggest to someone that it did something else.
Dec 27, 2017
0
#270
Elroch wrote:
SmyslovFan wrote:
Your condescension is noted.
It's not actually condescending to point out that someone should have no problem comprehending what AlphaZero does not do. I can see how anyone could be surprised that what it did do was so effective and that that might suggest to someone that it did something else.
lots of player playing QID because az did in my arena . i guess it created an impact. 
Dec 27, 2017
0
#271
mcris wrote:
Yes, and if it had opening book or let to play by itself in the opening, it would have lost 0 as Black.
Someone has already pointed out the interesting observation that Stockfish got solid positions out if the openings. The problems occurred later. Playing for a draw is not so easy when you have a 130 point Elo deficit.
Dec 28, 2017
0
#272
Elroch wrote:
mcris wrote:
Yes, and if it had opening book or let to play by itself in the opening, it would have lost 0 as Black.
Someone has already pointed out the interesting observation that Stockfish got solid positions out if the openings. The problems occurred later. Playing for a draw is not so easy when you have a 130 point Elo deficit.
That someone is a patzer compared to SF. SF never enters QID without opening book. Also does not play French defense like in the games with AZ.
Elroch wrote:
mcris wrote:
Yes, and if it had opening book or let to play by itself in the opening, it would have lost 0 as Black.
Someone has already pointed out the interesting observation that Stockfish got solid positions out if the openings. The problems occurred later. Playing for a draw is not so easy when you have a 130 point Elo deficit.
I am not so sure that Stockfish got solid positions out of the opening--its later play started with something out of the opening. [at least in the 10 games i saw]
Does Stockfish play for a draw or does it just play the best moves it can find?
Yes, he is.
Elroch wrote:
admkoz wrote:
Elroch wrote:
I am not sure what supposed co-incidence you are referring to, but AlphaZero has no way of explicitly remembering a position. Rather, it evaluates every position by finding the state of the neurons in its net, and it is fair to say the evaluation arising from those neuron states is based on the evaluations of other positions that has somewhat similar states for the neurons. (At the bottom level, this might mean the same pieces on certain squares, but at a higher levels it might include much more abstract concepts).
You said it dropped way off with less than 1/30th sec think time, which is 33ms, about = the 40ms it had during training. That's the coincidence.
As far as the part about neurons, I think that 'neuron' is a metaphor here and for me to discuss it usefully I would need to know what it actually represents in terms of software and how its state is driven by the state of the board. Probably beyond the scope of a blog discussion
Actually, neurons are easy to describe. They have a large number of real-valued inputs, combine these inputs with a linear function (determined by a parameter for each input) and then apply a fixed non-linear transformation (the activation function) before outputing a single real-valued output. The way to think of them is of combining lots of pieces of information to estimate some other piece of information. Other than the final outputs of a network, the neurons are not told by the designer what information to produce, they are just the inputs to other neurons.
That is fine, but not enough info for me to figure out how to get from A (randomness) to B (clobbering stockfish). Does that mean that a neuron, in terms of software, is basically a thread running some function? What are the large number of real valued inputs and how are those inputs derived from the actual board position? What is the initial linear function/non-linear transformation? How is it updated based on what happens in a random game? How much can we meaningfully discuss this before getting into things that only AZ's programmers know?
2.9 billion positions, number you threw out above, barely scratches the surface. 6 pieces alone can already achieve 50 billion positions. Somebody said it used "800 random playouts" to evaluate a position initially. It is hard to believe that 800 random playouts would achieve even one mate with K and Q vs K, let alone K and R. Each one would also take forever.
One major question I have is what actual breakthrough AZ's programmers made. Any, or did they just throw more hardware at it than anybody previously has?
yes magnus is patzer
I agree.
Dec 28, 2017
0
#278
Elroch wrote:
SmyslovFan wrote:
In what way am I incorrect, in the size of the number of games AlphaZero played before the match? Ok, it was several million games, not 240,000. That makes it more lopsided, not less.
This should not be difficult for you. AlphaZero only has an evaluation routine which it learnt from self-play and a search method which uses that evaluation routine. The games it has played are not remembered, only the evaluation routine, defined by a large number of learnt parameters.
I have read this and re-read this.
You are saying that the evaluations of these positions were saved, but not the positions themselves.
I really would like to see a citation for this extraordinary claim.
Dec 28, 2017
0
#279
For reference, here is the original paper, which repeatedly states that the engine learned from self-play. It gives charts explaining its analysis of openings. I can't find anywhere in the paper that the games it played are not remembered:
https://arxiv.org/pdf/1712.01815.pdf
Hobby player, you made a bunch of assumptions, most of which are currently fiction.
Even with a suboptimal program running without any database help, Stockfish only lost 4 games out of 50 as White. Naka believes that if it had database assistance it would have lost 0 as White.