Life In 19x19

macelee wrote:

Bill Spight wrote: But for analysis I would want at least 10K rollouts for each main option (all four in this case)

I was using my laptop which is not powerful enough to obtain 10k within reasonable amount of time. But even if we have 10k, I doubt it would pick up other locations like P4.

I would want 10K rollouts to have confidence in the difference in win rates. I expect that their accuracy depends in large part on the number of rollouts. √(1/10,000) = 1%, so with 10,000 rollouts for each option I would feel pretty confident that if option A has a win rate 3% better than the win rate for option B, then option B is a mistake, and if the difference is 2% it is probably a mistake. The easy way to do this is not to wait until the number of rollouts gets large enough, but to play each option and run them for 10K rollouts.

BTW, to use LZ, we seriously need to get a proper GPU. Last weekend, my son asked me to show him the inside of a computer after reading a computing book. I found a 12 year old computer in my garage and it actually still booted without problem. It has an old NVidia card, a card about 15 generations older than the current ones. When I compiled Leela Zero on that machine, to my surprise this very old card still managed to generate rollouts 4 times faster than my brand new laptop (using an Intel i7 CPU).

Back in the 1990s I asked Anders Kierulf why, what with Moore's Law, my new computer was slower than my computer from 10 years before. He just laughed.