Komodo Dragon : Knight Odds Handicap Match

[Rebel]

Given up on Lc0. While Arena dictates a -99.99 score for resignation the engine apparently resigns by itself.

That's strange; I found that recent Lc0 (largest net) was quite strong giving knight odds, either to me or to engines. I didn't see stupid moves. But I don't use Arena, perhaps there is some hidden problem there?

The 384x30 net played well. The Arena adjudication is just false. I will have to move Lc0 to cute. Meanwhile I am trying Ethereal.

[Nordlandia]

Do you plan to use MCTS mode for a event like this ?

[lkaufman]

Given up on Lc0. While Arena dictates a -99.99 score for resignation the engine apparently resigns by itself.

That's strange; I found that recent Lc0 (largest net) was quite strong giving knight odds, either to me or to engines. I didn't see stupid moves. But I don't use Arena, perhaps there is some hidden problem there?

The 384x30 net played well. The Arena adjudication is just false. I will have to move Lc0 to cute. Meanwhile I am trying Ethereal.

If you are using just one thread for all the CPU based engines, comparing those results to a GPU-based Lc0 result would be quite misleading. It would have to be compared to an MP result for the CPU engines to be meaningful. Of course if Lc0 doesn't score better on a GPU than the CPU engines do on one thread, no need to address this point.

[lkaufman]

Do you plan to use MCTS mode for a event like this ?

MCTS only scores better than standard mode in handicap games with humans; against engines it scores worse.

[Rebel]

Given up on Lc0. While Arena dictates a -99.99 score for resignation the engine apparently resigns by itself.

That's strange; I found that recent Lc0 (largest net) was quite strong giving knight odds, either to me or to engines. I didn't see stupid moves. But I don't use Arena, perhaps there is some hidden problem there?

The 384x30 net played well. The Arena adjudication is just false. I will have to move Lc0 to cute. Meanwhile I am trying Ethereal.

If you are using just one thread for all the CPU based engines, comparing those results to a GPU-based Lc0 result would be quite misleading. It would have to be compared to an MP result for the CPU engines to be meaningful. Of course if Lc0 doesn't score better on a GPU than the CPU engines do on one thread, no need to address this point.

Comparing apples with oranges is never fair, it is what CPU vs GPU is about, never perfect. However at the Gambit Rating List the 3th place of Lc0 looks fair.

[lkaufman]

Given up on Lc0. While Arena dictates a -99.99 score for resignation the engine apparently resigns by itself.

That's strange; I found that recent Lc0 (largest net) was quite strong giving knight odds, either to me or to engines. I didn't see stupid moves. But I don't use Arena, perhaps there is some hidden problem there?

The 384x30 net played well. The Arena adjudication is just false. I will have to move Lc0 to cute. Meanwhile I am trying Ethereal.

If you are using just one thread for all the CPU based engines, comparing those results to a GPU-based Lc0 result would be quite misleading. It would have to be compared to an MP result for the CPU engines to be meaningful. Of course if Lc0 doesn't score better on a GPU than the CPU engines do on one thread, no need to address this point.

Comparing apples with oranges is never fair, it is what CPU vs GPU is about, never perfect. However at the Gambit Rating List the 3th place of Lc0 looks fair.

Is that list comparing one thread on CPU to an RTX 1060 GPU for Lc0, or was Lc0 somehow slowed down to offset the hardware advantage? If it's a straight-up comparison, I'm surprised that Lc0 would do so poorly with the hardware advantage.

[Rebel]

Given up on Lc0. While Arena dictates a -99.99 score for resignation the engine apparently resigns by itself.

That's strange; I found that recent Lc0 (largest net) was quite strong giving knight odds, either to me or to engines. I didn't see stupid moves. But I don't use Arena, perhaps there is some hidden problem there?

The 384x30 net played well. The Arena adjudication is just false. I will have to move Lc0 to cute. Meanwhile I am trying Ethereal.

If you are using just one thread for all the CPU based engines, comparing those results to a GPU-based Lc0 result would be quite misleading. It would have to be compared to an MP result for the CPU engines to be meaningful. Of course if Lc0 doesn't score better on a GPU than the CPU engines do on one thread, no need to address this point.

Comparing apples with oranges is never fair, it is what CPU vs GPU is about, never perfect. However at the Gambit Rating List the 3th place of Lc0 looks fair.

Is that list comparing one thread on CPU to an RTX 1060 GPU for Lc0, or was Lc0 somehow slowed down to offset the hardware advantage? If it's a straight-up comparison, I'm surprised that Lc0 would do so poorly with the hardware advantage.

Yes, CPU vs GPU. Apparently the 1060 is not that fast, see the head-to-head statistic.

Code: Select all

 3) Lc0-v27        3429.1 :    800 (+307,=388,-105),  62.6 %

    vs.                   :  games (   +,   =,   -),   (%) :    Diff,    SD, CFS (%)
    sf13                  :    100 (   3,  58,  39),  32.0 :  -143.6,  10.4,    0.0
    Komodo-Dragon         :    100 (  13,  67,  20),  46.5 :   -51.5,  14.0,    0.0
    Komodo_14             :    100 (  43,  46,  11),  66.0 :   +97.7,  13.2,  100.0
    RubiChess_2.1         :    100 (  54,  40,   6),  74.0 :  +158.1,  11.9,  100.0
    SlowChess_2.5         :    100 (  51,  38,  11),  70.0 :  +165.5,  14.0,  100.0
    Igel_3.0.5            :    100 (  42,  50,   8),  67.0 :  +168.3,  13.7,  100.0
    Ethereal_12.75        :    100 (  52,  42,   6),  73.0 :  +174.6,  12.0,  100.0
    Igel_3.0.0            :    100 (  49,  47,   4),  72.5 :  +186.9,  10.8,  100.0

BTW, you got some competition from Ethereal.

Code: Select all

   Engine           Points Games 
1. Komodo Dragon 1   7.5    20
2. Stockfish 13      6.0    20
3. Ethereal 12.75    6.0    20 
4. Komodo Dragon 2   5.5    20