Progress on Blunder

[Rebel]

Blunder 7.3.0 for the GRL.

Games : 1600

Elo pool : 2500

http://rebel13.nl/a/grl.htm

[algerbrex]

Congrats Did you implement a pawn hash to store the results of the pawn structure evaluation?

Thanks

I actually did not get around to putting in a pawn hash. That mostly slipped my mind. Thanks for mentioning it again though, I'm going to make a note of it since it shouldn't be too hard to get working.

Understanding neural networks, at least the "normal" variety, isn't that difficult... the first thing (for me) would be to find out what the difference is between NNUE and a "normal" (matrix-calculation-based) NN is.

Yeah, even now I have a basic understanding of how a normal neural network works. And I think I have some idea of how to construct one for a chess engine right now if I wanted to.

But I figured I'd start with a NN first, and then once I understood those well, I would move to understand NNUE. Since I'm assuming one would need a good understanding of NN to fully grasp the advantages and design of NNUE.

[algerbrex]

Blunder 7.3.0 for the GRL.

Games : 1600

Elo pool : 2500

http://rebel13.nl/a/grl.htm

Thanks for the testing Ed. Interesting results so far.

[mvanthoor]

But I figured I'd start with a NN first, and then once I understood those well, I would move to understand NNUE. Since I'm assuming one would need a good understanding of NN to fully grasp the advantages and design of NNUE.

I assume you're not going to use some sort of neural network library, and write the network from scratch? (I probably will try that, but I won't even be thinking of that before Rustic is at least 2850 on a single thread and has Lazy SMP implemented. That'll be a -long- way in the future. I'd be very happy if I reach that stage before the end of the year.... 2022. But I doubt it.)

[Rebel]

Code: Select all

Gambit Rating List
Running      : gauntlet Blunder 7.3.0
Time Control : Time control : 40/120
Games        : 1600

Results from file gauntlet-blunder.pgn:

No. Name            Win Draw Loss Unf.  Score Games       %
-----------------------------------------------------------
  1 Blunder 7.3.0  +101  =52 -103  *19  127.0   256   49.6%
  2 CT800 1.43      +19   =5   -9   *1   21.5    33   65.2%
  3 Monolith 0.3    +18   =7   -8   *2   21.5    33   65.2%
  4 Loki 3.5.0      +12  =11  -10   *2   17.5    33   53.0%
  5 Galjoen 0.41.1  +15   =4  -11   *4   17.0    30   56.7%
  6 Foxsee 7.20.1   +12   =7  -12   *3   15.5    31   50.0%
  7 Jumbo 0.5.3     +12   =4  -16   *2   14.0    32   43.8%
  8 Orion 05         +9   =7  -16   *3   12.5    32   39.1%
  9 Jumbo 0.6.10     +6   =7  -19   *2    9.5    32   29.7%

Total Games:     275
White Wins:       96 (34.9%)
Black Wins:      108 (39.3%)
Draws:            52 (18.9%)
Unfinished:       19 (6.9%)

Estimated ratings for this elo 2500 pool

   # PLAYER            :  RATING  POINTS  PLAYED   (%)
   1 Monolith 0.3      :  2608.0    21.5      33    65
   2 CT800 1.43        :  2608.0    21.5      33    65
   3 Galjoen 0.41.1    :  2545.4    17.0      30    57
   4 Loki 3.5.0        :  2519.7    17.5      33    53
   5 Foxsee 7.20.1     :  2498.4    15.5      31    50
   6 Blunder 7.3.0     :  2498.4   127.0     256    50
   7 Jumbo 0.5.3       :  2454.3    14.0      32    44
   8 Orion 05          :  2420.5    12.5      32    39
   9 Jumbo 0.6.10      :  2347.3     9.5      32    30

Second time cute-chess crashed because Blunder crashed and so I leave it as is after 275 games.

Elo ~2500

[algerbrex]

But I figured I'd start with a NN first, and then once I understood those well, I would move to understand NNUE. Since I'm assuming one would need a good understanding of NN to fully grasp the advantages and design of NNUE.

I assume you're not going to use some sort of neural network library, and write the network from scratch? (I probably will try that, but I won't even be thinking of that before Rustic is at least 2850 on a single thread and has Lazy SMP implemented. That'll be a -long- way in the future. I'd be very happy if I reach that stage before the end of the year.... 2022. But I doubt it.)

Right, that’s the plan. And that’s still a long way off for me as well. But I figure it doesn’t hurt to start thinking about it now so I’ll be ready when the time does come.

[algerbrex]

Code: Select all

Gambit Rating List
Running      : gauntlet Blunder 7.3.0
Time Control : Time control : 40/120
Games        : 1600

Results from file gauntlet-blunder.pgn:

No. Name            Win Draw Loss Unf.  Score Games       %
-----------------------------------------------------------
  1 Blunder 7.3.0  +101  =52 -103  *19  127.0   256   49.6%
  2 CT800 1.43      +19   =5   -9   *1   21.5    33   65.2%
  3 Monolith 0.3    +18   =7   -8   *2   21.5    33   65.2%
  4 Loki 3.5.0      +12  =11  -10   *2   17.5    33   53.0%
  5 Galjoen 0.41.1  +15   =4  -11   *4   17.0    30   56.7%
  6 Foxsee 7.20.1   +12   =7  -12   *3   15.5    31   50.0%
  7 Jumbo 0.5.3     +12   =4  -16   *2   14.0    32   43.8%
  8 Orion 05         +9   =7  -16   *3   12.5    32   39.1%
  9 Jumbo 0.6.10     +6   =7  -19   *2    9.5    32   29.7%

Total Games:     275
White Wins:       96 (34.9%)
Black Wins:      108 (39.3%)
Draws:            52 (18.9%)
Unfinished:       19 (6.9%)

Estimated ratings for this elo 2500 pool

   # PLAYER            :  RATING  POINTS  PLAYED   (%)
   1 Monolith 0.3      :  2608.0    21.5      33    65
   2 CT800 1.43        :  2608.0    21.5      33    65
   3 Galjoen 0.41.1    :  2545.4    17.0      30    57
   4 Loki 3.5.0        :  2519.7    17.5      33    53
   5 Foxsee 7.20.1     :  2498.4    15.5      31    50
   6 Blunder 7.3.0     :  2498.4   127.0     256    50
   7 Jumbo 0.5.3       :  2454.3    14.0      32    44
   8 Orion 05          :  2420.5    12.5      32    39
   9 Jumbo 0.6.10      :  2347.3     9.5      32    30

Second time cute-chess crashed because Blunder crashed and so I leave it as is after 275 games.

Elo ~2500

Really? Odd. I haven’t had Blunder crash with cutechess in a long time, and not even in my recent testing. Would you happen to have anymore more info about the crashes?

[Rebel]

It says, Blunder engine terminated and then the whole of other running games are terminated by cute chess, then cute chess itself terminates.

[algerbrex]

It says, Blunder engine terminated and then the whole of other running games are terminated by cute chess, then cute chess itself terminates.

Hmmm, thanks. I'll have to do some investigating. Perhaps I'll have to release a hotfix, since no point in testing a buggy version that crashes. Time control was 40 moves/2 mins right?

Apologies for not testing more thoroughly.

[Rebel]

It says, Blunder engine terminated and then the whole of other running games are terminated by cute chess, then cute chess itself terminates.

Hmmm, thanks. I'll have to do some investigating. Perhaps I'll have to release a hotfix, since no point in testing a buggy version that crashes. Time control was 40 moves/2 mins right?

Apologies for not testing more thoroughly.

Yes, 40/120, 40m/2m. No problem, pretty common for brand new engines, I have restarted at 40/10, perhaps the PGN gives a hint.