I've discovered quite an interesting way to limit the playing strength of Leela 0.11.0 that I would like to share with you. I don't know the exact playing strength of Leela 0.11.0, but I think the program should be quite strong. So it finds the famous 37th move of the 2nd game between LeeSedol and AlphaGo practically in 1 second.
Now if I set the time to 1 minute per game, set the engine max level to 100 simulations, and set the power of the processor to 5% under battery power, then Leela plays on my laptop in a way that I can beat it as a DDK player.
Since the program - in contrast to e.g. MfoG - sticks to the computing time requirements, it comes to an average of 30 simulations for the main variant.
If I now experiment a bit with the CPU performance and the simulations, then a Leela can set it to my liking.
Maybe this info is useful for one or the other of you.
Weak Leela
-
Mike Novack
- Lives in sente
- Posts: 1045
- Joined: Mon Aug 09, 2010 9:36 am
- GD Posts: 0
- Been thanked: 182 times
Re: Weak Leela
I'm not sure what you are saying.
Only the top two levels of MFOG are using MCTS. Below that the levels are all using evaluation by "rules" and this is so fast, by comparison, that giving the program less time/CPU power will make little difference. But the program offers strength settings down to 18 kyu. And if using the "randomize" option, won't play quite the same each time in the same situation.
Too solid? Perhaps. No way I could judge that. Back in the days before MCTS came along (and I was weaker) I was overly "honte" myself. Not to the extent that when playing human opponents who played loosely I couldn't counter (perhaps because I did see what was honte or not). But I can see your point that neither of these programs would have given practice at doing that.
Only the top two levels of MFOG are using MCTS. Below that the levels are all using evaluation by "rules" and this is so fast, by comparison, that giving the program less time/CPU power will make little difference. But the program offers strength settings down to 18 kyu. And if using the "randomize" option, won't play quite the same each time in the same situation.
Too solid? Perhaps. No way I could judge that. Back in the days before MCTS came along (and I was weaker) I was overly "honte" myself. Not to the extent that when playing human opponents who played loosely I couldn't counter (perhaps because I did see what was honte or not). But I can see your point that neither of these programs would have given practice at doing that.