I can't believe that so many people don't get it!

[mhull]

But is that correct? There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning. Are there enough neurons to remember millions of these stats or is there more to it? In a human brain data storage and neurons are synonymous but in a computer brain neurons and data storage is not synonymous and yet if one wants to use the human brain function as a parable then memory and neurons would be talked about as they are synonymous. So yes the games as a separate entity have been discarded but the 'memory' of the games is stored--in a learn file.

It doesn't seem so much like a human brain which can remember an entire game "PGN". The NN system only remembers "principles". I think move legality and selection is run by the MCTS and the NN is consulted for positional scoring. As far as the NN is concerned, the game scores themselves are consigned to the memory hole.

[Michael Sherwin]

This is not what the paper describes.

Indeed. But we are not dealing with academics. We are dealing with a commercial company with a bad reputation. They have their mouth full of ethics but their actions are criminal, like the copying of books, like buying youtube while knowing its massive illegal copyrighted material and only God knows what they do with the data we publicly trust to the internet and thus to them, my name might as well be colored red after this post

Ranting aside, let's talk about what the paper doesn't reveal, the start positions of the 100 games. Those 50 start positions certainly can be learned as Mike described.

Thanks Ed! And another thought is that they might have started the main match right after the 4 hours of learning so the w,l,d,p data may have still been in memory for speed and not stored on a hard drive. That would make sense so they could say that the w,l,d,p data was part of the NN hiding the separation.

[hgm]

The 100 games all started from the normal start position. And all training games started also from that position. So yes, in a sense that is learning how to play good moves from the start position. But isn't that exactly what playing Chess is? Btw, from the way they describe it, I have no doubt that training on the standard opening position will make it do just as well in a match from any other start position. After all, the 100 match games for the largest part consisted out of positions AlphaZero could have never seen before, in these millions of training games. If it would not be able to do superb moves in positions it had never seen before, it would lose to Stockfish 100-0.

Even if you think they must be lying through their teeth because of their evil and greedy nature, anything you claim they would have been doing instead would still purely be a fabrication. You cannot possibly kow that. In that case the simplest explanation is that it is all a complete hoax.

But there also is this: in my judgement, what they describe is likely to work. What you think they have done instead: no way that could ever work...

[vvarkey]

There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning.

where is this evidence?

[Michael Sherwin]

But is that correct? There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning. Are there enough neurons to remember millions of these stats or is there more to it? In a human brain data storage and neurons are synonymous but in a computer brain neurons and data storage is not synonymous and yet if one wants to use the human brain function as a parable then memory and neurons would be talked about as they are synonymous. So yes the games as a separate entity have been discarded but the 'memory' of the games is stored--in a learn file.

It doesn't seem so much like a human brain which can remember an entire game "PGN". The NN system only remembers "principles". I think move legality and selection is run by the MCTS and the NN is consulted for positional scoring. As far as the NN is concerned, the game scores themselves are consigned to the memory hole.

As I wrote above they may have started the main match after the 4 hours of training thus keeping the MCTS in memory. They did keep actual w,l,d,p data for every move. The NN's job was to pick the best move. That is what the evidence is strongly suggesting. Once again there is no way they stored w,l,d,p information from millions of games in the NN without losing 99% of that information. People might be thinking in terms of pure NN and not a hybrid system also requiring some amount of stored data. For the type of reinforcement learning they describe that stored data would be around 20 to 30 gigabytes.

[Daniel Shawul]

Wow Micheal you are now at a stage of not implying but vividly saying i invented this shit.
What you don't get (or continually ignore) is your learning stuff is position and player specific -- and that is the elephant in the room.
If you learn something against rybka, it doesn't work against stockfish and other opponents .... That form of learning doesn't even work against the same opponent if it randomizes its moves. Ask yourself why Romi is not at Rybka's level even though it beat it with learning. When you come up with learning that works against every opponent, you may then claim you invented this kind of stuff.

The form of learning you use has been there forever and infact i may have had it at one time. Tournament organizers often forbid it as they consider it as a cheap trick.

The most general form of learning that is not opponent/posiiton specific is TD-lambda, which adjusts its evaluation weights. This one is a real algorithm.

To call monte-carlo tree search learning hugely demonstrates your ignorance with what AlphaGo is doing. The fact that MCTS stores its statistics tree in memory is not learning, you could have done the same with alpha-beta. Also traning a NN is same as training a handcrafted evaluation. What exactly are you accusing them of ?

Daniel

[mhull]

But is that correct? There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning. Are there enough neurons to remember millions of these stats or is there more to it? In a human brain data storage and neurons are synonymous but in a computer brain neurons and data storage is not synonymous and yet if one wants to use the human brain function as a parable then memory and neurons would be talked about as they are synonymous. So yes the games as a separate entity have been discarded but the 'memory' of the games is stored--in a learn file.

It doesn't seem so much like a human brain which can remember an entire game "PGN". The NN system only remembers "principles". I think move legality and selection is run by the MCTS and the NN is consulted for positional scoring. As far as the NN is concerned, the game scores themselves are consigned to the memory hole.

As I wrote above they may have started the main match after the 4 hours of training thus keeping the MCTS in memory. They did keep actual w,l,d,p data for every move. The NN's job was to pick the best move. That is what the evidence is strongly suggesting. Once again there is no way they stored w,l,d,p information from millions of games in the NN without losing 99% of that information. People might be thinking in terms of pure NN and not a hybrid system also requiring some amount of stored data. For the type of reinforcement learning they describe that stored data would be around 20 to 30 gigabytes.

But NNs don't store input data or output data. They use inputs/outputs to learn but keep neither. They only thing they keep is the values within the NN matrix which is not input data or output data.

[hgm]

As I wrote above they may have started the main match after the 4 hours of training thus keeping the MCTS in memory. They did keep actual w,l,d,p data for every move. The NN's job was to pick the best move. That is what the evidence is strongly suggesting. Once again there is no way they stored w,l,d,p information from millions of games in the NN without losing 99% of that information. People might be thinking in terms of pure NN and not a hybrid system also requiring some amount of stored data. For the type of reinforcement learning they describe that stored data would be around 20 to 30 gigabytes.

Sorry, but it seems you are making this up completely. What evidence? All of what you say is in direct contradiction with the paper. Even if you think they must be lying about everything they claim, how would that be evidence for what they did instead? How could you know more than what the paper states? Do you have a spy inside DeepMind?

[Michael Sherwin]

There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning.

where is this evidence?

My eyes are shot. This all I could find.

"Monte Carlo search does not use a tradition eval as we know it, so mobility, king safety etc. are irrelevant.

It uses a struct to hold info likes wins, losses, draws, win %, etc.,
then simply references accumulated data for the current position to select the move with the highest probability of winning."

The current position info has to be stored in memory somewhere!

[Daniel Shawul]

There is evidence that they keep a learn file with wins, losses, draws and a percentage chance of winning.

where is this evidence?

My eyes are shot. This all I could find.

"Monte Carlo search does not use a tradition eval as we know it, so mobility, king safety etc. are irrelevant.

It uses a struct to hold info likes wins, losses, draws, win %, etc.,
then simply references accumulated data for the current position to select the move with the highest probability of winning."

The current position info has to be stored in memory somewhere!

Wow you are so ignorant of what MCTS is doing it is not funny ...

Please stop trolling.