So Alpha Zero was a hoax?

[jhellis3]

lol.... Eyyyyyyyyyy.

[Milos]

So if someone advertises a product, e.g. a hand soap and says it removes 99.9% of bacteria and in reality in only removes 95% is that the same as if someone is advertising a hand soap that is actually a dishwashing liquid?

I thought you didn't like strawmans?

Nice of you to actually acknowledge Ronald's argument was a straw-man .

The statement that I don't know why you are not simply stating that DeepMind faked all their games was neither presented nor intended as an argument. My real point was that your claims are baseless. Also that was more a trivially true observation than an argument.

Your claim has exactly the same merit as my claim that Google falsely presented their training time, i.e. that they actually used 5000 TPUs in the same time for self-playing games, since there is actually zero evidence that prove the claim from the paper or disprove my claim.
So your point my friend is also baseless .

[Milos]

lol.... Eyyyyyyyyyy.

You're back to Hodor mode.

[jhellis3]

Ad hominem. You apparently do not suffer from cognitive dissonance. I have known a few humans like that. Nonetheless, quite remarkable.

[Milos]

Ad hominem. You apparently do not suffer from cognitive dissonance. I have known a few humans like that. Nonetheless, quite remarkable.

Yea sure, that's ad hominem, as much as your RIP is a death threat.
Chill out man, not having sense of humour or being proven wrong is not a tragedy. No need to throw a tantrum .

[noobpwnftw]

This argue about training time is going nowhere.

We can't really say that we spent 70 years on SF and they took 3 hours to surpass. After something gets invented and it just can do certain things way faster, like you can now have all perfect knowledge of 6 piece endgames or below in a weekend or less, so whoever invented this or made similar approaches surpassed mankind in a weekend because it is probably never possible otherwise?


The method, in sense of a chess program, is new in many ways and my primary interest is if it is ever going to work like they said, then the trade off between speed and evaluation quality can go to this far so we should probably try their MCTS implementation and use shallow a/b searches as a replacement to their NN, then to see if this thing works or not, or how fast the shallow a/b searches has to produce similar quality results from NN.
And we'd already know the reverse, putting a slow but better evaluation in a/b search doesn't work that well, this is many doubts' origin, I guess.

Then, in a chess form I pay zero attention to how reinforcement learning works in general, I take it for granted that it will work in whatsoever, sounds delusional but they got all the fanboys attached to it. So, yes, then what?

[jhellis3]

The RIP was a reference your own intellectual suicide not a threat of any kind YDB.

[noobpwnftw]

that a reinforcement learning-based approach is capable of competing with the traditional alpha-beta approach

So where is the search part? Is it not included in the preprint or something?

Your claim that chess is somehow different from Go, likewise, would have been much more convincing had it been made before the AlphaZero results were published.

Oh, really? There is it when I said about the differences and possible chess usage, now I throw this back to you, where were you when there is no AlphaZero to brag around?

Seriously dude, if you had made any contribution in chess programming in general, can you name some please, because from your history here I can only see a troll Based on this, looks like now you have finally found your Messiah.

[syzygy]

Your claim that chess is somehow different from Go, likewise, would have been much more convincing had it been made before the AlphaZero results were published.

Oh, really? There is it when I said about the differences and possible chess usage, now I throw this back to you, where were you when there is no AlphaZero to brag around?

Based on this, looks like you have finally found your Messiah?

Isn't David just saying that in May 2017, when you made that statement, it sounded completely convincing? What you wrote there is almost exactly what I was still saying in October and November 2017 after the AlphaGo Zero results had become known. Unfortunately for me though, the AlphaZero results rather strongly suggest that I got it all wrong.

Maybe I am giving Deep Mind too much credit, but I doubt it. We'll know more when others have succeeded or failed to replicate the results (I don't expect a Deep Mind paper to change many minds, regardless of what is in it).

[noobpwnftw]

Your claim that chess is somehow different from Go, likewise, would have been much more convincing had it been made before the AlphaZero results were published.

Oh, really? There is it when I said about the differences and possible chess usage, now I throw this back to you, where were you when there is no AlphaZero to brag around?

Based on this, looks like you have finally found your Messiah?

Isn't David just saying that in May 2017, when you made that statement, it sounded completely convincing? What you wrote there is almost exactly what I was still saying in October and November 2017 after the AlphaGo Zero results had become known. Unfortunately for me though, the AlphaZero results rather strongly suggest that I got it all wrong.

Maybe I am giving Deep Mind too much credit, but I doubt it. We'll know more when others have succeeded or failed to replicate the results (I don't expect a Deep Mind paper to change many minds, regardless of what is in it).

Well, the thing with him in the way I understand is a history digging about why I said so recently to support my doubts.

This guy just completely rejects any possibility that we can adapt certain parts of the DM research to make improvements rather than ditching everything we have and go for complete overhaul based on reinforcement learning and every aspect I said against NN is invalid because there is a preprint that in his belief had covered them all.

I don't think he have a clue about a/b is only a part of a chess program and such, since in his term is "we should ditch alpha-beta and use reinforcement learning approach", I guess proper wording would be "we should ditch handwritten evaluation and use reinforcement learning to train NNs that could do better then adjust search", if that makes him a better troll.

It still does not change my thoughts though, in a/b architecture I fail to see any other way to make good use of NNs. But that does not stop there, since we are unlikely to reproduce an NN implementation, what can be done is to adapt that search method and see which part contributed the most, is it the way it searches or the more accurate static evaluation.