Amazing results of LeelaKnightOdds-Dev

[Father]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

.. for me everything is very clear... LeelaQueenOdds, will soon be a 2500 FIDE elo level player. The LeelaQueenOdds Top 100 standings will hopefully be made up of Magnun Carlsen, Nakamura and the next best players in the world.

[Uri Blass]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

It seems that for some reason you downgraded lc0.

Decided to play another game and I do not remember it allow even trading so easily in the first games.

https://lichess.org/c8Avvw1A

I do not believe it is 2100 not even 2000 at 3+2 based on the last games.

[lkaufman]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

It seems that for some reason you downgraded lc0.

Decided to play another game and I do not remember it allow even trading so easily in the first games.

https://lichess.org/c8Avvw1A

I do not believe it is 2100 not even 2000 at 3+2 based on the last games.

That's a bit strange. The draw change only affects willingness to repeat, not relevant here. We did cut the max search in half, but our tests showed this should actually improve results on average, though of course it will hurt in some positions. Maybe the shorter search helps vs bots but hurts vs humans, who probably play the opening much better than similarly rated bots. There was one other technical change that shouldn't hurt. But I share your feeling that it got weaker somehow. It's not worth trying to address this, because it should be replaced very soon by a much better net, as I mentioned. But perhaps we should be wary of tests that tell us to use very small node limits.

[lkaufman]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

.. for me everything is very clear... LeelaQueenOdds, will soon be a 2500 FIDE elo level player. The LeelaQueenOdds Top 100 standings will hopefully be made up of Magnun Carlsen, Nakamura and the next best players in the world.

Well, that's a bit too optimistic in my opinion, but there is some reason to believe that LeelaQueenForKnight is already around that level if you are talking about 3'0" chess. Its latest results are hard to believe (I would suspect cheating if there was any imaginable way for the best odds engine in the world to cheat!). But the difference of a knight is huge.

[Uri Blass]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

It seems that for some reason you downgraded lc0.

Decided to play another game and I do not remember it allow even trading so easily in the first games.

https://lichess.org/c8Avvw1A

I do not believe it is 2100 not even 2000 at 3+2 based on the last games.

That's a bit strange. The draw change only affects willingness to repeat, not relevant here. We did cut the max search in half, but our tests showed this should actually improve results on average, though of course it will hurt in some positions. Maybe the shorter search helps vs bots but hurts vs humans, who probably play the opening much better than similarly rated bots. There was one other technical change that shouldn't hurt. But I share your feeling that it got weaker somehow. It's not worth trying to address this, because it should be replaced very soon by a much better net, as I mentioned. But perhaps we should be wary of tests that tell us to use very small node limits.

I thought maybe allowing trading pieces is to prevent repetitions in the search that I even do not think about
but it is not logical because I assume repetition should be a good result for leela at least when it is a full queen down and not knight down.

[lkaufman]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

You caught something here! Another major upgrade is planned, using a net actually trained for queen odds rather than one trained just for knight odds. Estimated gain was 150 elo. The actual Elo used for the current bot is 2100. The new numbers were posted a bit early, but are not being used until the switch actually happens, probably in a couple days. We want to be more sure of how much of an improvement it is and how to set parameters. I expect that the 2250 rating (at least) will be fully justified once it is released. I'll post here when it happens. The remaining bots may also be upgraded pretty soon. It's really remarkable how strong they are getting to be.

It seems that for some reason you downgraded lc0.

Decided to play another game and I do not remember it allow even trading so easily in the first games.

https://lichess.org/c8Avvw1A

I do not believe it is 2100 not even 2000 at 3+2 based on the last games.

That's a bit strange. The draw change only affects willingness to repeat, not relevant here. We did cut the max search in half, but our tests showed this should actually improve results on average, though of course it will hurt in some positions. Maybe the shorter search helps vs bots but hurts vs humans, who probably play the opening much better than similarly rated bots. There was one other technical change that shouldn't hurt. But I share your feeling that it got weaker somehow. It's not worth trying to address this, because it should be replaced very soon by a much better net, as I mentioned. But perhaps we should be wary of tests that tell us to use very small node limits.

I thought maybe allowing trading pieces is to prevent repetitions in the search that I even do not think about
but it is not logical because I assume repetition should be a good result for leela at least when it is a full queen down and not knight down.

Right, this draw stuff should not be relevant in the opening. Anyway I predict that after the net upgrade neither you nor "father" will be able to score 50% at your presently preferred time controls. I hope we don't get a situation where the leaderboard has no one scoring 50%, maybe we'll start to attract stronger players or players will choose longer time controls.

[Marcus91]

I can that I am surprised to see the following that mean increase in the rating of Leela relative to what I saw earlier:

LeelaQueenOdds as white 3+2 (Reference) 2249.6

If she is 2249.6 then I am better than 2300 at blitz based on my results against her but it is clearly not the case.

The leaderboard now displays ratings based on the current network. Ratings for the future network will be applied once the upgrade is complete. Leela's rating still needs final adjustments, so players' scores will adjust accordingly, but the relative differences between players will remain the same

[Uri Blass]

today not a good day against lc0 and I won only 3.5-2.5

Note that I did not do serious analysis to prepare but my conclusion that if I get a good position lc0 allow easily to trade pieces so serious preperation should be about getting a good position to learn from my positional mistakes in the opening.

Tactical mistakes when I already made some positional mistake are less important.

I had 2122.3 after 48 games when lc0 with white has 2099.6
I expect some small rating increase based on the 6 games.

total score so far is 35-19 for me(35-18 if I do not include one loss on time because of internet disconnection).

[Marcus91]

today not a good day against lc0 and I won only 3.5-2.5

Note that I did not do serious analysis to prepare but my conclusion that if I get a good position lc0 allow easily to trade pieces so serious preperation should be about getting a good position to learn from my positional mistakes in the opening.

Tactical mistakes when I already made some positional mistake are less important.

I had 2122.3 after 48 games when lc0 with white has 2099.6
I expect some small rating increase based on the 6 games.

total score so far is 35-19 for me(35-18 if I do not include one loss on time because of internet disconnection).

Get ready for the upgrade in the next few days, I can tell you It's going to be really tough!

[Uri Blass]

won additional 2 games inspite of not playing well and hope it is enough to get second place.

The leaderboard before these games was the following:
1 nn1133 2235.8 6166
2 Catecan 2133.0 272
3 Ajktulm 2130.4 127
4 uriblass 2129.9 54
5 Brunetticus 2105.5 245
6 KentGoulding 2079.8 234
7 Hissha 2057.1 54