Engine Tournament

[q30]

I had tested different LeelaZero versions with different weights details.
The rating of Go engines without using GPU is next:

Super level
1. LeelaZero;

Highest level
2. Leela;

High level
3. Hiratuka;

Middle level
4. Pachi_DCNN;
5&6. Ray & Pachi;
7. MoGo.

[q30]

LeelaZero 0.14 is a bit stronger (not as much, as in case of Phoenix), than 0.15 (details)...

[as0770]

LeelaZero 0.14 is a bit stronger (not as much, as in case of Phoenix), than 0.15 (details)...

My Russian is a little bit rusty. Do you seriously try to tell us v0.14 is better than v0.15 based on 6 games?

[q30]

v. 0.15 seriously is not better, than v. 0.14, and most likely weaker in case of Phoenix...

[as0770]

v. 0.15 seriously is not better, than v. 0.14, and most likely weaker in case of Phoenix...

Look at the changelog, it is no surprise that v0.15 is not stronger than v0.14. There are only changes for training games. Beside that it is ridiculous to draw conclusions from a 6 game match. There is a reason why the Leela Team does 400 game matches and requires a 55% score for the promotion of a network.

[q30]

In these results 0.14 wins 0.15 by black and by white. The proximity to 50 % may explained by very small time control...
And what about Phoenix?

[as0770]

In these results 0.14 wins 0.15 by black and by white. The proximity to 50 % may explained by very small time control...

51% winrate says nothing.

And if you play a match between two engines of the same strength the timecontrol doesn't matter. The chance of winning is 50%, no matter if you play with 1 visit, 100000 visits or with a rolling dice. You still don't understand mathematical basics.

[Uberdude]

Are you trying to find the affect of the Leela Zero engine version on the strength of the bot? Whilst it could make some difference if there are algorithmic changes or parameters tuned, most of the difference will come from the version of the weights used.

[as0770]

And what about Phoenix?

v0.14 doesn't even support the Phoenix network.

There was a hack of v0.14 that supported Phoenix networks, but it wasn't working correctly afaik.

[q30]

In these results 0.14 wins 0.15 by black and by white. The proximity to 50 % may explained by very small time control...

51% winrate says nothing.

And if you play a match between two engines of the same strength the timecontrol doesn't matter. The chance of winning is 50%, no matter if you play with 1 visit, 100000 visits or with a rolling dice. You still don't understand mathematical basics.

You still don't understand mathematical basics: if time control -> 0, than game of any (except pattern) engine -> to absolutely random and, respectively, the result of sparring -> 50%:50% nevertheless of engines strength...

[q30]

Are you trying to find the affect of the Leela Zero engine version on the strength of the bot? Whilst it could make some difference if there are algorithmic changes or parameters tuned, most of the difference will come from the version of the weights used.

It may depend on code productivity, i.e. it may was appended some time waste in new version...

[q30]

And what about Phoenix?

v0.14 doesn't even support the Phoenix network.

There was a hack of v0.14 that supported Phoenix networks, but it wasn't working correctly afaik.

See here. The file of weight is the same...

[as0770]

You still don't understand mathematical basics: if time control -> 0, than game of any (except pattern) engine -> to absolutely random and, respectively, the result of sparring -> 50%:50% nevertheless of engines strength...

You are riding a dead horse and make a fool out of yourself.

Leela with time=0 or visit=1 is far away from playing random. Try to win a game against it. And playing one engine against itself will result in a 50% winning chance regardless the time control. I don't know why I even have to discuss this basics again and again.

See here. The file of weight is the same...

Sometimes I wonder if you just want to fool us, but in the meantime I think you just don't understand English well enough to get the points. But in this case you should reconsider if you want to do this kind of discussions here.

Yes, the network is the same, but v.14 just can't handle it and the hack of v0.14 didn't handle it correctly.

[q30]

I don't know English well, so I'll try to answer on that I understood (and how I understood).
Yes, if AI Leela has pattern like play in case of time -> 0, than it will be not random, but will the same sequence (with the same weight file). So with the same weight file sparring of different Leela versions will -> 50%:50%. And playing one engine (version) against itself will result in a 50% winning chance regardless the time control only in case of using the same weight file...

I don't know, what You mean under 'v.14 just can't handle it and the hack of v0.14 didn't handle it correctly', but the versions 0.14 and 0.15 by link above (i.e. of Leela Phoenix) are playing with the weight file by the same link...

[as0770]

50% winning chance regardless the time control only in case of using the same weight file...

Or a weight file of similar strength. And once again: If one engine scores 55% in a 400 game match, there is still a chance of 1% that it is _not_ stronger.

I don't know, what You mean under 'v.14 just can't handle it and the hack of v0.14 didn't handle it correctly', but the versions 0.14 and 0.15 by link above (i.e. of Leela Phoenix) are playing with the weight file by the same link...

Your link points to LeelaZero_PhoenixGo v1 and v2 but you were talking about Leela Zero v0.14 and 0.15. LeelaZero 0.14 don't support Phoenix networks. Results of LeelaZero_PhoenixGo v1 vs. LeelaZero_PhoenixGo v2 with Phoeniox Networks don't say anything about the strength of LeelaZero v0.14 and 0.15. I wouldn't even trust results with Leela networks.