I haven't gone through the paper, clearly.

But it's also clear, at least to me, that holding a match at 1 min/move is not indicative of SF's real strength.

So why not, instead of 100 meaningless games, run 30 meaningful games at 3 min/move?

And let SF use the Syzygy tables too.

Because it's "standard operating procedure".

And because certain endgames are a "solved" problem.

Given that alpha-beta searches are basically brute-force anyways, why wouldn't tablebases be allowed? It helps the engine play stronger - and the challenge is for AlphaGo to be king of the hill without qualification.

.

But it's a bit of a joke, as the time scale goes from 10**-1 to 10**1 moves/second (it's a log scale).

Who cares what an engine's strength is at 0.1 seconds. And I'd sure like to see the upper limit extend beyond, let's be generous, 60 seconds.

I'd also have preferred two 50-game matches against two strong engines, then a 100-game match against only one. Or even three different engines.

But that's just me.

Then if it can learn in one minute what a human might derive in two weeks, for instance, AlphaZero's four hours would be equal to about 9 years, 24 hours a day, of human study.

I also read that it played, studied, 700,000 games before it played Stockfish. That would equate to a human playing over 1300 games a week for ten years and learning something from each game.

But so far I'm dubious on it being vastly superior to today's top programs. The two games I saw vs. Stockfish, it appeared Stockfish played more than a couple non-optimal moves in the opening or early mid-game getting itself into positions very uncharacteristic for a top computer.

My experience with SF8 is that it is prone to this kind of thing unless you allow it an extended amount of "crunch" time. I've previously posted examples of baffling irregularities.

<<...> But I'll guess we'll see if Alpha ever enters some computer v. computer tournament.>

This is indeed the "acid test".

If I read the graphs correctly, certain opening just aren't worth playing anymore, as AlphaGo gains experience.

Those openings include the following:

French, Sicilian, Spanish and C-K.

Well, AlphaZero does have knowledge, and while it doesn`t use traditional chess knowledge it has TONS of more knowledge than anything else in this world about chess.

Now if the 700,000 games number is correct, I find it hard to believe AlphaZero had no guidance.

Without guidance AlphaZero, would likely start learning with a move like 1. a3 or 1.a4. Would seem to make logical sense.

How many games would it take for AlphaZero to figure out 1.a3 is no good without either guidance or versus practice games against a strong engine that shortens the learning pattern?

It would seem by itself, without some sort of guidance, it would have to go through 1. a3 and all combinations versus all defences, then 1. a4, then 1. b3 , 1. b4 etc.

Maybe I'm wrong but that seem's like it would take more than 700,000 games to go through every possible opening combination and get through to an equal middle game. Forget about narrowing it down the ones that are best and then those proceeding to show advantages later in the game.

Especially if it's not programmed with some sort of evaluation function to shorten the learning process but has to play out all these games to the end to decide the merits. Given that there will likely be a number of 1.a3's versus a bad defense that might show it a reasonable contender.

Wow, if it played about 3 billion games in four hours. Then any guidance is probably not necessary. Now, I'm really looking forward to a tournament entry by AlphaOne.

In principle, you're right. But the effect on society of lawyers, doctors, etc losing their jobs would result in massive social instability, which would not be good news for anyone.>