Cardoso wrote:This "zero" approach really surprised me, starting from almost nothing except the game's rules. I wasn't expecting this advancement so soon, even though they used pretty powerful hardware.
One of the things that amazes me is that Alpha Zero isn't blind to deep tactics (taking into consideration it only runs at 80K positions per second).
Can someone try to explaine this?
I mean I can understand it plays well positionally, but also being tactical so strong is surprising to me.
The deep neural network connects the pieces on different squares to each other. They use 3x3 convolutions. This means that the next 8x8 layer's cells are connected to a 3x3 region (called "receptive field") in the previous region, and to a 5x5 region in the layer before etc. After only 4 layers, each cell is connected to every other cell in the original input layer. For AlphaGoZero they used no less than 80 layers. Then they also have many "feature maps" in parallel, so that they can learn different concepts related to piece-square combinations. Finally, they use the last 8 positions as input as well, so they also have a sense of ongoing maneuvers. All this is then being trained on the game result and the best move from the MC tree search.
Although the amount of resources required to train the millions of weights related to these neural networks is enormous, conceptually it is not surprising that pawn structure, king safety, mobility and even deep tactics can be detected from the last 8 positions.
Quite simple.
Scientific explanation others can give you. I can just tell you what it means.
Positional Play, Tactical Play.... these are just Terms used by humans and obsolete chess engines like Stockfish, Houdini and Komodo.
[b]The AlphaZero Program, after learning Chess in just a few Hours, has one and only one Goal, to Mate the Enemy King, NOTHING else ![/b]
What you see as Positional or Tactical Play, means NOTHING to this Monster Program. It wants to win and only win !
[b]Positional or Tactical Play, is only a means to an end, for this Program[/b].
It doesn't care what you call the type of Moves it makes.
[b]It just makes the Moves it needs to make, to achieve its goal of winning.[/b]
Simple but deadly.
[/quote]
You are totally right!
Humans can cogitate only in categories like positional, tactical, opened, closed, central, etc. Machines can not cogitate but they generate appropriate moves - that is all!
The deep neural network connects the pieces on different squares to each other. They use 3x3 convolutions. This means that the next 8x8 layer's cells are connected to a 3x3 region (called "receptive field") in the previous region, and to a 5x5 region in the layer before etc. After only 4 layers, each cell is connected to every other cell in the original input layer. For AlphaGoZero they used no less than 80 layers. Then they also have many "feature maps" in parallel, so that they can learn different concepts related to piece-square combinations. Finally, they use the last 8 positions as input as well, so they also have a sense of ongoing maneuvers. All this is then being trained on the game result and the best move from the MC tree search.
Although the amount of resources required to train the millions of weights related to these neural networks is enormous, conceptually it is not surprising that pawn structure, king safety, mobility and even deep tactics can be detected from the last 8 positions.
[/quote]
Where read you the above?
In your view the 80 layers enough to treat all of chess problems?
shrapnel wrote:
Quite simple.
Scientific explanation others can give you. I can just tell you what it means.
Positional Play, Tactical Play.... these are just Terms used by humans and obsolete chess engines like Stockfish, Houdini and Komodo. The AlphaZero Program, after learning Chess in just a few Hours, has one and only one Goal, to Mate the Enemy King, NOTHING else !
What you see as Positional or Tactical Play, means NOTHING to this Monster Program. It wants to win and only win ! Positional or Tactical Play, is only a means to an end, for this Program.
It doesn't care what you call the type of Moves it makes. It just makes the Moves it needs to make, to achieve its goal of winning.
Simple but deadly.
We can say that Alpha Zero is a nominalist chess engine?
We can say that Alpha Zero is a nominalist chess engine?
[/quote]
AlphaZero is not a chess engine.
It is a modern computer system dedicated to solve complicated problems like playing go, chess, shogi etc.
Systems like AlphaZero can be used for solving problems of scientific, economic, military, social behavior (even if political issues).....
Dann Corbit wrote:Comparing AlphaZero to Stockfish is like comparing an airplane to a motorboat.
Both can be used to cross a sea, but they are using completely different techniques to get there.
SF won't run on tensor chips.
AlphaZero won't run on ordinary CPUs.
The approach is very different.
The horsepower of AlphaZero is enormous. But that does not diminish the outcome. If we gave SF a trillion cores, the last 99.999999999 % don't help anyway,
Everything is comparable, there is no such thing in the Universe that is not comparable.
Alpha is a 2800 engine, making use of tremendous power.
Any guess why are they stressing the word 'artificial intelligence', when their intelligence/software on single core is at 2800, so immensely below any of the middle-tier engines?
Why don't they simply stress the hardware point?
Dann Corbit wrote:Comparing AlphaZero to Stockfish is like comparing an airplane to a motorboat.
Both can be used to cross a sea, but they are using completely different techniques to get there.
SF won't run on tensor chips.
AlphaZero won't run on ordinary CPUs.
The approach is very different.
The horsepower of AlphaZero is enormous. But that does not diminish the outcome. If we gave SF a trillion cores, the last 99.999999999 % don't help anyway,
Everything is comparable, there is no such thing in the Universe that is not comparable.
Alpha is a 2800 engine, making use of tremendous power.
2800 engines are not going to beat stockfish8 at LTC even with 100:1 hardware advantage.
100:1 is less than 7 doubling and you get less than 50 elo at long time control by doubling so less they will be less than 3200 elo even if you give them 100:1 hardware advantage.