pilgrimdan wrote:
I do not have trouble accepting that Stockfish lost the 100 game match...
the 'trouble' I am having is why do it under non-optimal conditions...
can you explain that...
What is optimal depends on what you're optimising under what constraints.
One constraint they would have had is time allocation on the machine they ran their matches on. Presumably there were other teams running stuff on the machine. It's not like they had a dedicated machine for running chess matches for an arbitrary amount of time, they would have needed to allocate computer time and use that time as best as they could. Running fixed time per move matches probably makes that easier, and using a small hash for Stockfish means you don't penalise it too much if a game is terminated because the time slice is up and you restart the game later.
That's pure speculation, but practical concerns like that probably had an effect on the choice of match conditions. It's not interesting enough to mention it in the paper though. These are minor details.
Leto wrote:
I would think they'd release all the games when their paper gets published, that will be a while. What we've seen so far is just a preview of their paper.
Probably.
It's worth considering too that 99% of readers are not interested in a ton of different games. They want statistics and a few illustrative examples. Which is what we got.
There's nothing strange about this at all.
syzygy wrote:You wouldn't even need to fire up the engines, you could just let some GUI select the moves from the Cerebellum book.
Cerebellum is like A0, it always chooses the same moves, from both sides, and always ends up in this position:
They may both be deterministic (provided AlphaZero's search time is kept constant), but the similarities end there.
syzygy wrote:But as it is there is no real indication that it was just opening luck. AlphaZero also won the 12 matches played from other opening positions.
What we want to know is if, say, Alpha Zero was tested under CCRL conditions, if it would be at the top of the list, and if so, by how much would its advantage be to the second engine. Are there tests we can make to approximate this data?
CCRL approximately fixes the hardware, and AlphaZero does not run on that hardware.
syzygy wrote:CCRL approximately fixes the hardware, and AlphaZero does not run on that hardware.
Ofc it does don't spread misinformation.
You can run any kind of NN with TensorFlow framework on CPU only.
It would have few hundred evals per second and be like 500+ Elo weaker but that's their problem, that would be totally fair comparison regarding CCRL rules.
Evert wrote:One constraint they would have had is time allocation on the machine they ran their matches on. Presumably there were other teams running stuff on the machine. It's not like they had a dedicated machine for running chess matches for an arbitrary amount of time, they would have needed to allocate computer time and use that time as best as they could. Running fixed time per move matches probably makes that easier, and using a small hash for Stockfish means you don't penalise it too much if a game is terminated because the time slice is up and you restart the game later.
That's pure speculation, but practical concerns like that probably had an effect on the choice of match conditions. It's not interesting enough to mention it in the paper though. These are minor details.
That such an amazing crap that is beyond belief. Google used 5000+ TPUs for training, and had to share 4 TPUs for running the match with other teams, really????
You obviously think Google is some kind of a startup that is running from the garage or shed and not a multi-billion global advertising monopolists.
syzygy wrote:CCRL approximately fixes the hardware, and AlphaZero does not run on that hardware.
Ofc it does don't spread misinformation.
You can run any kind of NN with TensorFlow framework on CPU only.
It would have few hundred evals per second and be like 500+ Elo weaker but that's their problem, that would be totally fair comparison regarding CCRL rules.
Yeah. And you can emulate an x86 machine on a 6502, and then run Stockfish on that!
hgm wrote:Yeah. And you can emulate an x86 machine on a 6502, and then run Stockfish on that!
No you can't because of the memory, address space, etc.
And your comment is totally irrelevant, i.e. trolling.
If you actually pulled a head out of your arse you'd see that in Gary's LC0 translation of L0 for 16x smaller NN than in A0 case even the strongest GPU is no match for, e.g. 16 core x86 machine.