Komodo 11.3.1

lkaufman · Post by **lkaufman** » Tue Mar 06, 2018 8:17 pm

Komodo 11.3.1 is now available at komodochess.com. It is a bugfix of 11.3. The bug only affects X moves in Y minutes type time controls, Komodo simply wouldn't run with such time controls. 11.3.1 fixes that. If you only use game/x or game/x plus increment time controls, no need to download this version. It is free for anyone who bought or is entitled to 11.3. No need for anyone to restart any rating tests other than 40/20' type of tests.
On another topic, the new Skill levels, they were intended to be run on single thread (except full strength level 20). However the code does not require this, and it turns out that running on multiple threads raises the strength part way towards the next higher level. So rather than treating it as a bug to fix, we are calling it a feature. Since the levels are roughly 150 elo apart, many users might want an intermediate level between (for example) level 10 and 11, and so they can select level 10 but set threads to a higher number (up to the max threads on your system, typically 8 on i7 quadcore machines). As a rough guide I estimate 2 threads raises elo by 20, 4 threads by 40, 8 threads by 60, 16 threads by 80 elo. The exact amount depends on the level. Note that elo gains for multiple threads are much larger than this for full strength (level 20).

Frank Quisinsky · Post by **Frank Quisinsky** » Tue Mar 06, 2018 11:47 pm

Hi Larry,

please have a look here:

One of the FEOBOS statistics ...
Based on 41.614 analysis from A00-E99 for each of the engines you can see in the table.

And now, please have a look under:
Main variant with one move only ...

That isn't good for analysis!!
I await from a TOP programm in 100% of cases the complete main line.

Quantity of 0,00 positions can be smaller.
8% is to high!

In more as 98% with speculative 3-fold in main line.
Great results by Andscacs for an example!

In my opinion is such an update much more important as 10 Elo more. So I am loud thinking, hope that is OK for you.

All in all ...
Thanks again for your update service.
I like the Komodo development and wish you and your partner Mark good luck for the future.

Best
Frank

mjlef · Post by **mjlef** » Wed Mar 07, 2018 12:38 am

Frank Quisinsky wrote:Hi Larry,

please have a look here:

One of the FEOBOS statistics ...
Based on 41.614 analysis from A00-E99 for each of the engines you can see in the table.

And now, please have a look under:
Main variant with one move only ...

That isn't good for analysis!!
I await from a TOP programm in 100% of cases the complete main line.

Quantity of 0,00 positions can be smaller.
8% is to high!

In more as 98% with speculative 3-fold in main line.
Great results by Andscacs for an example!

In my opinion is such an update much more important as 10 Elo more. So I am loud thinking, hope that is OK for you.

All in all ...
Thanks again for your update service.
I like the Komodo development and wish you and your partner Mark good luck for the future.

Best
Frank

Are these results from games, or fro analysis? Was UCI_AnalyseMode or "go infinite" used? If UCI_AnalyseMode is on, then Komodo will eventually output a full PV. When playing games, if Komodo has used its maximum time, then there might not be a PV available to display. Anyway, you should have this on if full length PVs are important for you.

Dann Corbit · Post by **Dann Corbit** » Wed Mar 07, 2018 1:15 am

mjlef wrote:
Frank Quisinsky wrote:Hi Larry,

please have a look here:

One of the FEOBOS statistics ...
Based on 41.614 analysis from A00-E99 for each of the engines you can see in the table.

And now, please have a look under:
Main variant with one move only ...

That isn't good for analysis!!
I await from a TOP programm in 100% of cases the complete main line.

Quantity of 0,00 positions can be smaller.
8% is to high!

In more as 98% with speculative 3-fold in main line.
Great results by Andscacs for an example!

In my opinion is such an update much more important as 10 Elo more. So I am loud thinking, hope that is OK for you.

All in all ...
Thanks again for your update service.
I like the Komodo development and wish you and your partner Mark good luck for the future.

Best
Frank
Are these results from games, or fro analysis? Was UCI_AnalyseMode or "go infinite" used? If UCI_AnalyseMode is on, then Komodo will eventually output a full PV. When playing games, if Komodo has used its maximum time, then there might not be a PV available to display. Anyway, you should have this on if full length PVs are important for you.

In my experience, both SF and H are worse offenders when it comes to K for truncation of pvs.

I actually have a code patch that prevents SF from the horrid sewing machine motion it gets when the crazed fail high/fail low sequence starts to happen.
Well sort of. It still goes nuts, but I do not print out the truncated pvs.

Dann Corbit · Post by **Dann Corbit** » Wed Mar 07, 2018 1:18 am

Code: Select all


                // When failing high/low give some update &#40;without cluttering
                // the UI&#41; before a re-search.
                if (   mainThread
                        && multiPV == 1
                        && &#40;bestValue <= alpha || bestValue >= beta&#41;
                        && &#40;Time.elapsed&#40;) > 3000&#41; && Options&#91;"Show Fail High and Fail Low"&#93;)
                    sync_cout << UCI&#58;&#58;pv&#40;rootPos, rootDepth, alpha, beta&#41; << sync_endl;

Uri Blass · Post by **Uri Blass** » Wed Mar 07, 2018 1:56 am

Frank Quisinsky wrote:Hi Larry,

please have a look here:

One of the FEOBOS statistics ...
Based on 41.614 analysis from A00-E99 for each of the engines you can see in the table.

And now, please have a look under:
Main variant with one move only ...

That isn't good for analysis!!
I await from a TOP programm in 100% of cases the complete main line.

Quantity of 0,00 positions can be smaller.
8% is to high!

In more as 98% with speculative 3-fold in main line.
Great results by Andscacs for an example!

In my opinion is such an update much more important as 10 Elo more. So I am loud thinking, hope that is OK for you.

All in all ...
Thanks again for your update service.
I like the Komodo development and wish you and your partner Mark good luck for the future.

Best
Frank

Andscacs show wrong analysis based on watching TCEC.
It often print a full pv but play a move that is not the first move in the pv.

Do you think that it is better than having a short pv of only one move but the move it play?

Dann Corbit · Post by **Dann Corbit** » Wed Mar 07, 2018 2:00 am

Uri Blass wrote:
Frank Quisinsky wrote:Hi Larry,

please have a look here:

One of the FEOBOS statistics ...
Based on 41.614 analysis from A00-E99 for each of the engines you can see in the table.

And now, please have a look under:
Main variant with one move only ...

That isn't good for analysis!!
I await from a TOP programm in 100% of cases the complete main line.

Quantity of 0,00 positions can be smaller.
8% is to high!

In more as 98% with speculative 3-fold in main line.
Great results by Andscacs for an example!

In my opinion is such an update much more important as 10 Elo more. So I am loud thinking, hope that is OK for you.

All in all ...
Thanks again for your update service.
I like the Komodo development and wish you and your partner Mark good luck for the future.

Best
Frank
Andscacs show wrong analysis based on watching TCEC.
It often print a full pv but play a move that is not the first move in the pv.

Do you think that it is better than having a short pv of only one move but the move it play?

I think by definition the move it plays is a one node pv.

Any engine can fail high or low and play a different move when time runs out than the last pv node.

Maybe Andscacs does it more often than others for some reason.

Frank Quisinsky · Post by **Frank Quisinsky** » Wed Mar 07, 2018 5:09 am

Hi Mark,

1. 41.614 end positions to openings A00-E99, 3 moves after ECO codes formed.

2. Shredder Classic GUI, 60 seconds per position on Intel i7, 10 threads (6 cores).

End positions can be found in FEOBOS basic download.
Analysis of 13 engines (all 41.614 positions) can be found under FEOBOS analysis.

Hint:
Stockfish and Houdini have here much more problems, can be see in graphics, not important, both points ...

1. 0.00 output (to high)
2. complete line of analysis

... should be fixed!

Best
Frank

Frank Quisinsky · Post by **Frank Quisinsky** » Wed Mar 07, 2018 5:14 am

Hi Uri,

Andscacs have an own UCI parameter for such things. All works wonderful for myself. Furthermore, Andscacs and Booot ... for an example ... found the most best moves, 3 moves after ECO codes formed.

So you can see in KECR that Komodo and Stockfish aren't the best engines for opening analysis. Houdini here the number 1, Booot on 2, Andscacs on 3.

But differents are small.
Much more interesting this stats to each of the ECO codes. Very easy to see to which ECO code which engines comes with the best analysis.

We can many learn with FEOBOS and all the Excel statistics. Things, normally we can't see.

Best
Frank

Komodo 11.3.1

Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1

Re: Komodo 11.3.1