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

[kyotosato]

The purpose of this thread is first to clearly answer the question of how to use Phoenix go in sabaki ui. secondly, is to give some sample games of Phoenix go vs. leelaZ, and Phoenix go vs. deepZen 7 (commercial version.)
1. using Phoenix go in sabaki UI,
requirements are 64 bit cpu, and gpu (may be comfirmed by observing task manager if gpu is employed when Phoenix go or leelaZ is/are running.) And assuming that you have installed sabaki correctly and know how to operate it.
the zipped Phoenix Go file you should download is PhoenixGo-win-x64-gpu-v1.zip. unzip the file to a directory, say, myphoenix.
VERY IMPORTANT: the weight text file MUST be placed in the unzipped Phoenix Go's directory

in sabaki manu,
Engines -> Manage Engines -> Add
*(Unnamed Engine) ----- replaced by ---- phoenixGo_10_seconds ---- or whatever name you like.
*Path ----- replaced by ---- c:\somedirectories\myphoenix ---- or whatever directory your Phoenix go unzipped to.
*no arguments ---- replaced by ----- "-g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000"
*Initial commands (;-saparated) ---- replaced by ----- "time_settings 0 11 1;"
the 11 time setting will give you 10 seconds thinking time, likewise, 6 will give you 5, 21 will give 20, and so on.

2. presumably you have already installed leelaZ with the same parameters as above,so that you may play Phoenix go against leelaZ, and in my another thread shows you how to use DeepZen 7 in sabaki. now i will report supported by very limited sampling games,and the result:
Phoenix Go (takes white) always defeats leelaZ and DeepZen7 commercial under same time settings. (i am mostly interested in 20 seconds time setting. my cpu is i5.) BTW, Phoenix go weight file and leelaZ weight file are interchangeable, meaning that you may use leelaZ's weight file in Phoenix Go instead of using its own. this is noted in another Phoenix Go thread, and i tried and confirmed that it is true.

next post i will show the actual game sample(s).

[kyotosato]

game setting
PhoenixGo: v.1
arguments: -g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000
initial commands: time_settings 0 21 1;
Leela-zero v.0.15:
arguments: -g --noponder -t 4 -w leelaz-model-swa-32-256000.txt -p 500000
initial commands: time_settings 0 21 1;
(note: the weight text file is latest available dated 6/29/2018, and the 21 is for constant 20 seconds thinking time)
computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home

Score Estimation: White won by 46.5 point at time Black resigned (employs DeepZen7's score estimation function).
Observations:
1. PhoenixGo was never able to overcome winrate 50% to offset 6.5 Komi.
2. the default resignation threshold may be set up on winrate dropping to below 5%, this has to check out both LeelaZ and PhoenixGo specs.

[Tryss]

computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home

You're missing the most important part : what GPU did you use? Or is it without GPU? How many playouts do you get in 20s?

[kyotosato]

@Tryss
1. As my first post said you might check task manager to see if the GPU is running for both leelaZ and PhoenixGo.
"gpu (may be comfirmed by observing task manager if gpu is employed when Phoenix go or leelaZ is/are running.)"
And I did observed both processes in task manager employed GPU. As for the Asus box. i just don't have time to check out its GPU specification, however it's GPU spec may easily be checked out by googling this particular brand and model. i just got the PC a few days ago. I will check it out soon.

2. the average playouts per move is around 160 to 190 for leelaZ. PhoenixGo did not listing playouts reading. my guess is about the same range.

3. i just got a game record with PhoenixGo taking black and resigned using above 20s setting. LeelaZ won. I will post the game soon.

[kyotosato]

game setting
PhoenixGo: v.1
arguments: -g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000
initial commands: time_settings 0 21 1;
Leela-zero v.0.15:
arguments: -g --noponder -t 4 -w leelaz-model-swa-32-256000.txt -p 500000
initial commands: time_settings 0 21 1;
(note: the weight text file is latest available dated 6/29/2018, and the 21 is for constant 20 seconds thinking time)
computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home


Score Estimation: White won by 46.5 point at time Black resigned (employs DeepZen7's score estimation function).
Observations:
1. PhoenixGo was never able to overcome winrate 50% to offset 6.5 Komi.
2. the default resignation threshold may be set up on winrate dropping to below 5%, this has to check out both LeelaZ and PhoenixGo specs.

[abcd_z]

2. game rule setting of 6.5 Komi may not results in a fair game. my next post will switch position by letting PhoenixGo takes black with 6.5 komi.

Why would you do this in the first place? Both LZ and Phoenixgo were trained with a komi of 7.5.

[kyotosato]

@abcd_z:
"Why would you do this in the first place? Both LZ and Phoenixgo were trained with a komi of 7.5.
"
i don't get it, please elaborate your point. is that you mean the game should set 7.5 komi to be a fair game?

[kyotosato]

game setting
PhoenixGo: v.1
arguments: -g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000
initial commands: time_settings 0 21 1;
zen7: uses gtp4zen-0.3.2 for sabaki
arguments: -z 7 -t 4 -T 21 -s 200000

computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home


PhoenixGo (white) resigned because of ladder problem. but before PhoenixGo got into ladder problem, Zen7's winrate was constantly dropping from 47% to a point of no return.

[Tryss]

2. the average playouts per move is around 160 to 190 for leelaZ. PhoenixGo did not listing playouts reading. my guess is about the same range.

It's very (very) low. Both bots are crippled at such playouts numbers.

It's kinda like if you organised a match between two pro at a 1s/move speed

i don't get it, please elaborate your point. is that you mean the game should set 7.5 komi to be a fair game?

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?

[kyotosato]

@Tryss:
so, the ave. playouts should be at least how high to reach ama. 9 dan? and how high for pro level, say, pro 2 dan? a ballpark figure.

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.

[kyotosato]

game setting
PhoenixGo: v.1
arguments: -g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000
initial commands: time_settings 0 21 1;
zen7: uses gtp4zen-0.3.2 for sabaki
arguments: -z 7 -t 4 -T 21 -s 200000

computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home



PhoenixGo (white) won 0.5 point

[Bill Spight]

So Black 309 is the game losing move?

[kyotosato]

@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%.

[kyotosato]

game setting
PhoenixGo: v.1
arguments: -g --noponder -t 4 -w PhoenixGo_v1.txt -p 500000
initial commands: time_settings 0 21 1;
zen7: uses gtp4zen-0.3.2 for sabaki
arguments: -z 7 -t 4 -T 21 -s 200000

computer: Asus i5 ; cpu: Intel® Core™ i5-8250U CPU@ 1.60GHz 1.80GHz (64 bit)
ram: 4.00GB usable; OS: windows 10 Home



zen 7 (white) won. a weird game at the end. both bots behaved strangely. Zen7 (white) casted the first pass at move 270, the game should be over, and PhoenixGo (Black) won 9.5. Something wrong with PhoenixGo. At the end i manually stop the game.

[Uberdude]

@Tryss:
so, the ave. playouts should be at least how high to reach ama. 9 dan? and how high for pro level, say, pro 2 dan? a ballpark figure.

I don't know about PhoenixGo, but for LeelaZero I'd recommend at least 3000 playouts per move. On just a few hundred it has good intuition and judgement (9d+) but can make silly (kyu) reading mistakes, particularly with ladders, semeais, shortage of liberties etc. For a serious match I'd recommend at least 10k, 50k would be good.