Using Leela

[Bill]

I downloaded the Leela go program two days ago, and I feel likes it's
already improved my play. I play a game against it, and then review it
with the "Move Probabilities" turn on. Clearly, the moves suggested are
not always correct, but consideration of them definitely improves my
play. Someone (like me) with knowledge of the fundamental strategies of
the game, but who hasn't integrated them thoroughly, and is willing to
think, can make sense of the suggestions.

My question is: What are the "Move Probabilities" (or, where do they
come from)?

By the way, I noticed that the "analysis" part of the program runs my
CPU (i7-4790K) at 100% (getting the cores into the high temp range ~73C.
), so I reduced my use of that feature since I haven't set up my
computer for those extremes. This is the first program I've run that
uses more than 1 core of my cpu--and use it, it does! ; )

[w***@gmail.com]

Post by Bill
I downloaded the Leela go program two days ago, and I feel likes it's
already improved my play. I play a game against it, and then review it
with the "Move Probabilities" turn on. Clearly, the moves suggested are
not always correct, but consideration of them definitely improves my
play. Someone (like me) with knowledge of the fundamental strategies of
the game, but who hasn't integrated them thoroughly, and is willing to
think, can make sense of the suggestions.
My question is: What are the "Move Probabilities" (or, where do they
come from)?
By the way, I noticed that the "analysis" part of the program runs my
CPU (i7-4790K) at 100% (getting the cores into the high temp range ~73C.
), so I reduced my use of that feature since I haven't set up my
computer for those extremes. This is the first program I've run that
uses more than 1 core of my cpu--and use it, it does! ; )

Hi Bill,

in case you (or anyone else) is still interested in an answer:

1) the move probabilities come from the neural network that Leela uses to filter out the most likely moves that are worth exploring deeper. They come from millions of games (I think from KGS) that have been fed into the neural net, that way it learns from human games what a human would play in a given situation. If you're more interested in this topic I suggest you google a bit around for "alphago policy network"... it is a complicated subject when it comes to the details of how it works.

2) If you are running windows and have some current GPU (graphics card), you may try to use the OPENCL version of Leela. Basically, this switches the main load of analysis computations to the GPU, leaving your CPU idle. It works nicely for me on Linux, but it was a little messy to get it going. On windows this should be no big deal, just install the latest driver for your graphics card. I agree it is a little annoying that Leela will just hog 100% CPU power, making the machine very unresponsive. Even the "next move" click will be a challenge ;-)

[w***@gmail.com]

One thing I forgot: you can switch from "probable moves" (F3) to "best moves" (F4). That way, the moves will change as the analysis progresses, and very often the most likely moves will not be the best moves.

[Bill]

Post by w***@gmail.com
One thing I forgot: you can switch from "probable moves" (F3) to "best moves" (F4). That way, the moves will change as the analysis progresses, and very often the most likely moves will not be the best moves.

Hi! Thank you for this suggestion (and your other ones, including about
powering Leela by the GPU). I will investigate as time permits! Besides
still working on my Go game, I'm learning the Python programming language.

Bill

[m***@gmail.com]

Try this one: play Go with AI online http://joort.com/