PhoenixGo Installation and vs. Leela zero, vs. DeepZen 7

[kyotosato]

@Uberdude:

thanks for the info, i see if i could prolong each move time settings to increase playouts, and see if my i5 box will get a decent playouts within a tolerable time limit. for zen 7, to reach 10k per move is very fast and tolerable unless there is a race.

[Bill Spight]

@Bill Spight:
i don't understand why Zen7 continue until to move 309 to send "pass" move to end the game. the quesion is the move 309 of black is necessary or not? but winrate shows 99. % by PhoenixGo constantly for last 10 white moves. And Zen 7 analysis function shows move 309 is necessary if not place the move , black winrate would drop to 39%.

If Black 309 passes, what does White do?

Edit:


Edit2: Added some more fun variations.

[Tryss]

@Uberdude:

thanks for the info, i see if i could prolong each move time settings to increase playouts, and see if my i5 box will get a decent playouts within a tolerable time limit. for zen 7, to reach 10k per move is very fast and tolerable unless there is a race.

If you get 190 playouts in 20s, you'll need over 5 minutes to get 3000 playouts.

And I repeat again, but the CPU is mostly irrelevant for LZ, the GPU is what's important (A current top CPU is probably at least 10x slower than an average 5 year old GPU)

[kyotosato]

@Bill Spight:
move 309 was not necessary after all. did you manually get exhaustive variation? or you employed leelaZ to get them? just asking. pass or not pass decision design seems quite different from one to another.

[Bill Spight]

@Bill Spight:
move 309 was not necessary after all. did you manually get exhaustive variation? or you employed leelaZ to get them? just asking. pass or not pass decision design seems quite different from one to another.

I sensed that Black 309 was unnecessary right away. I did see the flaws in Black's position, but did not do any exhaustive reading before my original post. Afterwards I decided to make an SGF file. I did not see the main line that I showed all the way to the end, but such an underneath connection is quite familiar to me. I added variations because I think that it's a good idea to do so. One thing that I had not sensed was a variation I did not include, namely after White has captured the single stone, instead of playing atari, simply connect at G-15. White can cut off the Black stones to the left, but cannot win because of eye vs. no eye. That's one to remember. Actually, I have added some more variations, just for fun.

[kyotosato]

@Tryss:
"the CPU is mostly irrelevant for LZ, the GPU is what's important,"
is it a good assessment for go game as a whole, or for that matter any NP/PSpace/Expspace complete problem if applicable? moreover, DeepZen 7 does not employ GPU at all. or i should say full machine AI (zero) depending on GPU to get ahead of human experience dependent AI.

[kyotosato]

@Bill Spight:
As far as I know endgame problem differs from other stages of the game mathematically. thanks for showing all or almost all variation, my gut feeling tells me 309 is not necessary, and i will take chance to win.

[Tryss]

@Tryss:
"the CPU is mostly irrelevant for LZ, the GPU is what's important,"
is it a good assessment for go game as a whole, or for that matter any NP/PSpace/Expspace complete problem if applicable? moreover, DeepZen 7 does not employ GPU at all. or i should say full machine AI (zero) depending on GPU to get ahead of human experience dependent AI.

The most computationally expensive (by far) of LZ algorithm is the evalation of moves by the neural network. And GPU are much much more efficient at this than CPU.

Why? A GPU is built to do a huge numbers of (simple) parallels computations : it needs to compute the color of each pixel of the screen each frame.

But huge numbers of simple computations in parallel are exactly the kind of stuff you need to do to evaluate a neural network.

A modern CPU has between 4 and 8 independant cores, a modern GPU has over a thousand (but they are much simplier and not as flexible as the CPU cores).

[abcd_z]

about 7.5 or 6.5 komi problem, understood. still the sample game results are valid so long as white won. but it is beyond the point that i am questioning the komi setting of either 7.5, or 6.5 or even 5.5 are too high for a fair game setting.

on a second thought,
"He means that both bots play will play like there's 7.5 komi no matter what komi you set, so why using 6.5 komi?"
this is not a correct assessment while applying a real game situation. the value function to optimize next move is based on the current points advantage/disadvantage, which means the 6.5 or 7.5 komi setting does matter to generate each moves. The training is training patterns to store in neural nets, as far as i understand regardless the setting to 7.5 or 6.5 komi setting. or isn't it? i am not sure now.

quoting a poster from a different thread:

In the Champions League there are 3 Leela networks. They don't care about the komi, all results of them are due to resignation. So playing with a komi of 7.5 would have lead to the same results, right?

Not necessarily. In fact, Leela Zero should only lose by resignation, if she lose 0.5pt, then there is definitively something wrong. Same goes for AQ, Leela and Ray.

But let's take and example, LZ (black) vs Zen (white), komi=6.5pt
20 moves before the end of the game, black (LZ) is about to win by 0.5pt, seems good for LZ. But because LZ is trained for 7.5pt komi, she believes she will loose by 0.5 (her value network gives a 23% win rate), so she starts desperate attempts at complicating the game, play weird end game, lose points here and there, then her win rate drops bellow 20% and finally she resigns.

Zen was suppose to loose, but he ends up winning, Leela Zero resigns from a completely lost position.

[Uberdude]

And on CPU vs GPU::

My new laptop comes with a NVidia card. It is the so-called Optimus setting: the integrated GPU with the Intel processor handles day to day jobs, and the NVidia card can be used to do more intensive jobs.

...

But I have no idea how to use the GPU on this machine to run Leela Zero. Has anyone tried this before?

I was troubled by the poor performance of using the Intel CPU to run LZ. I knew it would be poor. But I was quite shocked when I tried LZ with a 10-year-old NVidia card on a very old desktop PC - the old machine was about 4 times faster.