New engine: Stash

[Scally]

Hi,

Your new Stash 9.0.1 is compiling fine on my Raspberry Pi.

Thanks very much

Al.

[mhouppin]

Current rating list (reference Stash 3.0):

Code: Select all

   # PLAYER       :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Stash 9.0    :     419    204    16.5      20  82.5      80   15    3    2  15.0
   2 Stash 8.1    :     340    185    15.0      20  75.0      61   11    8    1  40.0
   3 Stash 8.0    :     316    178    14.5      20  72.5      86   10    9    1  45.0
   4 Stash 8.2    :     227    167    12.5      20  62.5      69    7   11    2  55.0
   5 Stash 7.0    :     186    171    11.5      20  57.5      90    6   11    3  55.0
   6 Stash 2.0    :      84    160     9.0      20  45.0      78    5    8    7  40.0
   7 Stash 5.0    :      22    156     7.5      20  37.5      50    4    7    9  35.0
   8 Stash 6.0    :      22    167     7.5      20  37.5      60    3    9    8  45.0
   9 Stash 3.0    :       0   ----     7.0      20  35.0      50    3    8    9  40.0
  10 Stash 4.0    :       0    157     7.0      20  35.0      99    3    8    9  40.0
  11 Stash 1.0    :    -277    219     2.0      20  10.0     ---    0    4   16  20.0

White advantage = 62.54 +/- 28.47
Draw rate (equal opponents) = 63.41 % +/- 9.17

[mvanthoor]

Nice progression Can those engines be downloaded anywhere, or are they tagged in the repository?

You said you estimated version 8 at around 1450 ELO; and that version is almost 600 ELO stronger than version 1. That would mean that a minimal chess engine that just 'learned to play' can be expected to perform at around 850 ELO. There are random movers that are stronger than that :p

Did you test Stash against any engines that are on the CCRL or CEGT lists?

[Alayan]

There are random movers that are stronger than that :p

Random movers are thousands of elo below a non-buggy basic AB engine. Their CCRL ratings are extremely inflated because they play against buggy engines that throw away games for no reason whatsoever, and do not obey the assumption of the elo scale.

[mvanthoor]

Yes, Mr. Müller said the same thing in another thread. To be honest, I hadn't even thought about bugs in engines with regard to search or evaluation. I hope to avoid them by adding only very small new functionality to the engine, one tiny step at a time, and testing this thoroughly before adding something else. It does make development (in the beginning) very slow though.

Still, seeing an engine starting out improve something like 600 ELO is motivating I wonder if I can download or compile the different versions of Stash from somewhere.

[mhouppin]

Yes, Mr. Müller said the same thing in another thread. To be honest, I hadn't even thought about bugs in engines with regard to search or evaluation. I hope to avoid them by adding only very small new functionality to the engine, one tiny step at a time, and testing this thoroughly before adding something else. It does make development (in the beginning) very slow though.

There is indeed a "search bug" in the first version of Stash (which was corrected in 2.0, that explains the Elo difference) : I implemented incorrectly the iterative deepening, and allow for "aborted search" return values to overwrite the last iteration, leading to misinterpretations of the best move by the engine.

As the version 2.0 has also bugs with en passant captures, I'm using Stash 3.0 as a reference during testing, because I cannot trust older versions ^^

Still, seeing an engine starting out improve something like 600 ELO is motivating I wonder if I can download or compile the different versions of Stash from somewhere.

All major releases are tagged (v1, v2... v9), so on downloading the repository you can simply checkout on the version you want

[mhouppin]

Hi there, version 10.0 is out, with a new component: Quiescence Search !

Tests are still running to show the Elo improvement, but this is the first version to get consistent victories against the low rated engines of the CCRL (such as Dreamer 0.3.0). I'm now working on adding a Transposition Table.

[Gabor Szots]

Please tell me your country. The first set of tests with 9.0.1 is nearing its end and we put the author's nationality on our pages if it is available.

[mhouppin]

Please tell me your country. The first set of tests with 9.0.1 is nearing its end and we put the author's nationality on our pages if it is available.

I'm French. I'm currently doing assets with compiled binaries for principal 64-bit architectures so that it becomes possible to use them without a compiler.
Thanks a lot for the tests ! ^^

[PK]

Random movers are thousands of elo below a non-buggy basic AB engine.

Not only that, there is at least 600 Elo difference between a proper random mover and one that knows it pays off not to lose a queen.