Request for testing Scorpio 2.8.4 MCTS

[Daniel Shawul]

This is a request for testing the MCTS version of Scorpio available here

Scorpio 2.8.4 MCTS

I have been fixing bugs thanks to TCEC!

It should be similar level as default Scorpio at fast time control, but maybe 100 elos weaker on long time control (i don't know). I would really like to see it in rating lists like CEGT/CCRL !

Fun fact: It doesn't require a freaking monster hardware to perform well , just 1 CPU core should be enough for it .

Daniel

[Graham Banks]

Hi Daniel,

does it require a nice graphics card?

Graham.

[Daniel Shawul]

One poor little CPU core should be enough, however, it can run on two or more too.

Cheers

Daniel

[Dann Corbit]

I am running a test, but it may not be interesting to you.

I am running a contest with:
Scorpio 2.8.4 MCTS 1 thread
Scorpio 2.8.4 MCTS 2 threads
Scorpio 2.8.4 MCTS 4 threads
Scorpio 2.8.4 MCTS 8 threads

To see how it scales.

I saw some oddities so far, like 1thread getting the upper hand over 8threads.

[Dann Corbit]

So far, you can see that 1 thread seems to be punching above his weight:

Code: Select all

  Program                Elo    +   -   Games   Score   Av.Op.  Draws
1 Scorpio-384-mcts:1th : 3114  453 362     5    70.0 %   2966   20.0 %
2 Scorpio-384-mcts:8th : 3006  367 367     5    50.0 %   3006   20.0 %
3 Scorpio-384-mcts:4th : 2948  279 300     5    40.0 %   3018   40.0 %
4 Scorpio-384-mcts:2th : 2932  279 300     5    40.0 %   3003   40.0 %

OK, I know it is very early. But I get a strange vibe watching the games.

[Daniel Shawul]

So far, you can see that 1 thread seems to be punching above his weight:

Code: Select all

  Program                Elo    +   -   Games   Score   Av.Op.  Draws
1 Scorpio-384-mcts:1th : 3114  453 362     5    70.0 %   2966   20.0 %
2 Scorpio-384-mcts:8th : 3006  367 367     5    50.0 %   3006   20.0 %
3 Scorpio-384-mcts:4th : 2948  279 300     5    40.0 %   3018   40.0 %
4 Scorpio-384-mcts:2th : 2932  279 300     5    40.0 %   3003   40.0 %

OK, I know it is very early. But I get a strange vibe watching the games.

I guess I haven't fixed all the parallel bugs yet. TCEC uncovered so many of them.

The single core version, on the other hand, is well tested.

BTW can you send me your MSVC project file, or did you use my mingw build?

I wasn't sure it will work after I added all the InterlockedExchange stuff.

Daniel

[Dann Corbit]

So far, you can see that 1 thread seems to be punching above his weight:

Code: Select all

  Program                Elo    +   -   Games   Score   Av.Op.  Draws
1 Scorpio-384-mcts:1th : 3114  453 362     5    70.0 %   2966   20.0 %
2 Scorpio-384-mcts:8th : 3006  367 367     5    50.0 %   3006   20.0 %
3 Scorpio-384-mcts:4th : 2948  279 300     5    40.0 %   3018   40.0 %
4 Scorpio-384-mcts:2th : 2932  279 300     5    40.0 %   3003   40.0 %

OK, I know it is very early. But I get a strange vibe watching the games.

I guess I haven't fixed all the parallel bugs yet. TCEC uncovered so many of them.

The single core version, on the other hand, is well tested.

BTW can you send me your MSVC project file, or did you use my mingw build?

I wasn't sure it will work after I added all the InterlockedExchange stuff.

Daniel

The link you posted did not have source code, only binaries.

Currently:

Code: Select all

  Program                Elo    +   -   Games   Score   Av.Op.  Draws 
1 Scorpio-384-mcts:2th : 3053  125 122    20    60.0 %   2983   40.0 %
2 Scorpio-384-mcts:8th : 3001  134 134    20    50.0 %   3001   30.0 %
3 Scorpio-384-mcts:4th : 2974  110 112    20    45.0 %   3009   50.0 %
4 Scorpio-384-mcts:1th : 2972  110 112    20    45.0 %   3007   50.0 %

I can make you a project if you give me a link to the proper code.

[Dann Corbit]

Watching the games, there are often shocking swings in eval.
I do not know if this is normal for MCTS, but it does not look like watching alpha-beta games.

Of course, the game count is small, so it might all be coincidence

[Daniel Shawul]

Watching the games, there are often shocking swings in eval.
I do not know if this is normal for MCTS, but it does not look like watching alpha-beta games.

Of course, the game count is small, so it might all be coincidence

This is a combined MCTS+alphabeta engine that is somewhat difficult to parallelize.
Right now it is doing a PV-split kind of algorithm for parallelization where I have to wait the first move on the PV is searched fully before parallelization. Without that it were giving really wild bugs.

If you are interested in the scaling of only MCTS, I suggest you try setting "frac_alphabeta 0" in scorpio.ini. That is also very easy to parallelize and probably has no smp bugs.

The code is here https://github.com/dshawul/Scorpio/commits/master

Daniel