The PSTs of Carnivor

[Mike Sherwin]

Code: Select all

#PST from Carnivor by Mike Sherwin| Source: http://talkchess.com/forum3/viewtopic.php?f=7&t=76831&p=886615

Pawn 100
  0|  0|  0|  0|  0|  0|  0|  0
 40| 60| 60| 80| 80| 60| 60| 40
 20| 40| 40| 60| 60| 40| 40| 20
 10| 16| 12| 30| 30| 12|  8|  6
  8| 14| 16| 20| 20|  8|  6|  4
 12|  8|  0|-10|-10|  0|  4|  6
  4|  4|-24|-30|-30|  8|  8|  4
  0|  0|  0|  0|  0|  0|  0|  0

Knight 290   
  0|  6| 12| 12| 12| 12|  6|  0
  6| 12| 24| 36| 36| 24| 12|  6
 12| 24| 36| 40| 40| 36| 24| 12
 12| 24| 36| 42| 42| 36| 24| 12
 12| 24| 36| 40| 40| 36| 24| 12
 12| 24| 30| 36| 36| 38| 24| 12
  6| 12| 24| 30| 26| 24| 12|  6
  0|-10| 12| 12| 12| 12|-10|  0

Bishop 330
 6 |  0|  0|  0|  0|  0|  0|  6
 0 | 18| 20| 20| 20| 20| 18|  0
 0 | 12| 30| 24| 24| 30| 12|  0
 0 | 30| 24| 32| 32| 24| 30|  0
 16| 12| 32| 32| 32| 32| 12| 16
 0 | 20| 30| 12| 12| 30| 20|  0
 10| 30| 12| 12| 12| 12| 30| 10
 6 |  0|  0|  0|  0|  0|  0|  6

Rook 500
 22| 22| 22| 22| 22| 22| 22| 22
 24| 24| 24| 24| 24| 24| 24| 24
  6|  8| 10| 12| 12| 10|  8|  6
  4|  6|  8| 10| 10|  8|  6|  4
  4|  6|  8| 10| 10|  8|  6|  4
  6|  8| 10| 12| 12| 10|  8|  6
  8| 10| 10| 22| 22| 10| 10|  8
  8| 10| 12| 20| 20| 12| 10|  8

Queen 900
 18| 18| 18| 18| 18| 18| 18| 18
 18| 20| 20| 20| 20| 20| 20| 18
 16| 18| 20| 20| 20| 20| 18| 16
 16| 18| 20| 20| 20| 20| 18| 16
 14| 16| 18| 18| 18| 18| 16| 14
 14| 16| 18| 18| 18| 18| 16| 14
 14| 16| 16| 16| 16| 16| 16| 14
 14| 14| 14| 14| 14| 14| 14| 14

King 66666			
-10|-10|-10|-10|-10|-10|-10|-10
-10|  0|  0|  0|  0|  0|  0|-10
-10|  0| 20| 20| 20| 20|  0|-10
-10|  0| 20| 40| 40| 20|  0|-10
-10|  0| 20| 20| 20| 20|  0|-10
-20|-20|-20|-20|-20|-20|-20|-20
-30|-30|-30|-30|-30|-30|-30|-30
 20| 60| 20|-40|  0| 20| 60| 20

The POV of my files are as the white player sees a chess board. I'm running a few games now against my default values. And because I've been testing all the other PSTs against Rustic too I'm also playing yours against Rustic Alpha 1 now. Let's see if Carnivore's can beat texel-tuned Sunfish's! (which perform best against Rustic)

The kings probably do need an endgame PST. I can try reducing the bonus for g1/8, h1/8 and b1/8, a1/8 at some point. Like I said Car was a toy engine. I just like the style of play and once the kings are fixed it should do okay. But, probably not with the kings like they are now.

[lithander]

The kings probably do need an endgame PST. I can try reducing the bonus for g1/8, h1/8 and b1/8, a1/8 at some point. Like I said Car was a toy engine. I just like the style of play and once the kings are fixed it should do okay. But, probably not with the kings like they are now.

This might explain why Car's PSTs didn't do so well in my tests.

Code: Select all

Score of MinimalChess 0.3 Carnivor vs MinimalChess 0.3: 126 - 134 - 177  [0.491] 437
Elo difference: -6.4 +/- 25.1, LOS: 31.0 %, DrawRatio: 40.5 %

Self play strength was about equal but the draw ratio was really high because both versions often were in disagreement in their evaluation. When both thought they were losing they both accepted a draw by 3-ply-move-repetition. (I might bump the contempt value again above zero, I really don't like to see my engine play lame moves like that in a balanced position)

Against Rustic it played much worse with Carnivor's PSTs compared to e.g. Sunfish's.

Code: Select all

Score of MinimalChess 0.3 Carnivor vs Rustic: 217 - 576 - 207  [0.321] 1000
Elo difference: -130.5 +/- 20.1, LOS: 0.0 %, DrawRatio: 20.7 %

Code: Select all

Score of MinimalChess 0.3 Sunfish vs Rustic: 321 - 488 - 191  [0.416] 1000
Elo difference: -58.6 +/- 19.6, LOS: 0.0 %, DrawRatio: 19.1 %

But if you give me a new KING table I'll run these tests again. I got a lot of idle cores during the day! And playing chess is at least more worthwhile than mining cryptocoins

[Mike Sherwin]

The kings probably do need an endgame PST. I can try reducing the bonus for g1/8, h1/8 and b1/8, a1/8 at some point. Like I said Car was a toy engine. I just like the style of play and once the kings are fixed it should do okay. But, probably not with the kings like they are now.

This might explain why Car's PSTs didn't do so well in my tests.

Code: Select all

Score of MinimalChess 0.3 Carnivor vs MinimalChess 0.3: 126 - 134 - 177  [0.491] 437
Elo difference: -6.4 +/- 25.1, LOS: 31.0 %, DrawRatio: 40.5 %

Self play strength was about equal but the draw ratio was really high because both versions often were in disagreement in their evaluation. When both thought they were losing they both accepted a draw by 3-ply-move-repetition. (I might bump the contempt value again above zero, I really don't like to see my engine play lame moves like that in a balanced position)

Against Rustic it played much worse with Carnivor's PSTs compared to e.g. Sunfish's.

Code: Select all

Score of MinimalChess 0.3 Carnivor vs Rustic: 217 - 576 - 207  [0.321] 1000
Elo difference: -130.5 +/- 20.1, LOS: 0.0 %, DrawRatio: 20.7 %

Code: Select all

Score of MinimalChess 0.3 Sunfish vs Rustic: 321 - 488 - 191  [0.416] 1000
Elo difference: -58.6 +/- 19.6, LOS: 0.0 %, DrawRatio: 19.1 %

But if you give me a new KING table I'll run these tests again. I got a lot of idle cores during the day! And playing chess is at least more worthwhile than mining cryptocoins

Thanks Thomas, I appreciate you taking the time to do this. I will work on the king table. This morning I downloaded Sunfish and I read that it doesn't distinguish between middlegame and endgame so having a look at the Sunfish king table its g1 bonus is 40 instead of CAR's 60 and there is a path of low resistance for the king to come into the center via g2, g4, f5, f6 then it absolutely hates e6 (-67) for the king but loves d6 (+44). That part makes no sense at all unless it somehow acts synergistically with the other tables. This confirms the notion that I need to reduce g1 to 40 which is exactly what I was planning to try first anyway. I'll be back!

[lithander]

For what it's worth here's a KING table I use for the PSTs that are no sunfish and would otherwise require a distinct endgame table like rustic and the "simple evaluation" one.

Code: Select all

King 66666			
-10| 10|  10| -20| -20|  10| 10| -10
-20|  0|  20|  20|  20|  20|  0| -20
-30|  0|   0|  10|   0|   0|  0| -30
-30|  0|   0|   0|   0|   0|  0| -30
-30|-10|   0|   0|   0|   0|-10| -30
-20|-20| -20| -30| -30| -20|-20| -20
  0|  0| -10| -30| -30| -10|  0|   0
 10| 20|  10| -10| -10|   0| 20|  10

What I wanted the KING to do is to prefer castling and stay there but once he left the corner for whatever reason he'd get "pulled" into the center and there he'd find a few 20s on the 7th rank so that he wouldn't go back to the castling corners when the pressure there subsided but could actually help with promotion instead. And I gave a penalty to the corners. In self play of MMC 0.3 this modified simple evaluation did win more games against sunfish's tuned PSTs than it lost. But against Rustic the sunfish version did much better and Marcel observed MMC 0.3 not converting won endgames into actual victories. So I guess it didn't turn out so great after all!

This is why I'll want to give tapered eval ala PeSTO a try over the following weekend. But I'm also very curious what you can cook up because sunfish somehow managed to encode some endgame knowledge into it's PST through the automatic tuning process and I really wonder if a good human chess player could too. (I can't obviously, but I'm only rated 1000 on chess.com)

[mvanthoor]

This is why I'll want to give tapered eval ala PeSTO a try over the following weekend. But I'm also very curious what you can cook up because sunfish somehow managed to encode some endgame knowledge into it's PST through the automatic tuning process and I really wonder if a good human chess player could too. (I can't obviously, but I'm only rated 1000 on chess.com)

One of the reasons your PST above can't convert wins is because the king gets stuck on the 7th rank. There is not enough incentive to leave it, so you can't mate with K+R vs. K because you can't see 15-20 moves ahead (yet). What you need, is a table where the center is 0, and all of the other squares are negative. Thus, the king wants to be in the center if he is not there already. Then, the other king needs to be driven to the edge.

For other pieces, you can make somewhat generalized tables, like I did; and a tuner can do it better. For the king, you can't really encode both the opening and end game table into one, because they contradict. A tuner can do it, by jacking with the values for the other pieces, and by encoding other positional knowledge in the PST's. That's the reason why the tuned tables look so weird; they encode knowledge that the evaluation function itself doesn't have yet. (Thus, you need to retune all variables in the evaluation function as soon as you add an evaluation term.)

The trick is to make each step away from the center negative, and make the margin larger each time. So, the first step away is -20 (20 down from previous), but the next is -50 (30 down from previous) and the edge is -90 (40 down from previous). Thus, the king in the center is going to assist with mate, because he can now step out of the center: doing so is negative, but when it drives the other king to the edge, it will be MORE negative for that king.

If king A is in the center, his PST value is 0.
If king B is one step out of the center, his value is -20.

King A thus has a gross value of +20, compared to king B.

Now if a rook drives king B one step closer to the edge, king B's value becomes -50.
Now King A must follow, out of the center, or mate will never occur. When HE steps out of the center, his value becomes -20.

But... compared to King B, he is now at +30.

Thus, King A will step out of the center, because for King A, the evaluation will go from +20 better to +30 better than king B.

Essentially, this is already a tiny version of a tapered evaluation, with a very hard switch from 0% endgame to 100% endgame, for one piece only.

[Mike Sherwin]

For what it's worth here's a KING table I use for the PSTs that are no sunfish and would otherwise require a distinct endgame table like rustic and the "simple evaluation" one.

Code: Select all

King 66666			
-10| 10|  10| -20| -20|  10| 10| -10
-20|  0|  20|  20|  20|  20|  0| -20
-30|  0|   0|  10|   0|   0|  0| -30
-30|  0|   0|   0|   0|   0|  0| -30
-30|-10|   0|   0|   0|   0|-10| -30
-20|-20| -20| -30| -30| -20|-20| -20
  0|  0| -10| -30| -30| -10|  0|   0
 10| 20|  10| -10| -10|   0| 20|  10

What I wanted the KING to do is to prefer castling and stay there but once he left the corner for whatever reason he'd get "pulled" into the center and there he'd find a few 20s on the 7th rank so that he wouldn't go back to the castling corners when the pressure there subsided but could actually help with promotion instead. And I gave a penalty to the corners. In self play of MMC 0.3 this modified simple evaluation did win more games against sunfish's tuned PSTs than it lost. But against Rustic the sunfish version did much better and Marcel observed MMC 0.3 not converting won endgames into actual victories. So I guess it didn't turn out so great after all!

This is why I'll want to give tapered eval ala PeSTO a try over the following weekend. But I'm also very curious what you can cook up because sunfish somehow managed to encode some endgame knowledge into it's PST through the automatic tuning process and I really wonder if a good human chess player could too. (I can't obviously, but I'm only rated 1000 on chess.com)

Just my opinion but 20s as high values on the 7th rank is too close to the eighth rank if the engine has any depth at all. However, it seems like your test are at very fast time controls so maybe, idk, the depth your engine reaches is not enough to see that if the king is in the center that it can go to the 7th to help a pawn promote. I think what I'm trying to say is that if the engine searches deep enough the high scores should be in the center so it can head anywhere on the board as needed. Why do the PSTs have to be 100% static? Can't we just work from copies and modify the copies according to the game phase? Or do like I did in RomiChess just compute the PSTs at the root. So just copy the static tables at the root and then adjust the values with mobility for a minimal engine and anything that can be thought of for a more advanced engine.

Have you tested against TSCP1.81? If so can you post the results for the various PSTs. I recommend TSCP because it does have a standard and decent evaluation function. It might result in better test data because it would not be as likely to fall into the trap a > b > c > a when testing PSTs because all static PSTs will be strong in some positions and weak in others.

As for me and Bric right now we are on this road and a huge tree has fallen across the road and we can't get past it. It's just the strangest bug I have ever had to squash. The white king when it reaches a point that it is clear that the game is lost makes a beeline for a1 and when it gets there it commits suicide by capturing itself. Now I just saw it from the black side all the way from g8 to a1. It happens about 1 in 20 games. And only in Arena and not in the debugger where it plays faultlessly. I'm so frustrated right now that I just have to walk away for a moment and think what to try next. The following king PST seems better. It did convert a game that it would not have otherwise.

[pgn][Event "Fruit"]
[Site "MASTER"]
[Date "2021.03.11"]
[Round "1"]
[White "Bricabrac"]
[Black "Tscp181"]
[Result "1-0"]
[BlackElo "2200"]
[ECO "A67"]
[Opening "Benoni"]
[Time "10:09:18"]
[Variation "Four Pawns, Taimanov, 8...Nbd7"]
[WhiteElo "2200"]
[TimeControl "0+8"]
[Termination "adjudication"]
[PlyCount "172"]
[WhiteType "human"]
[BlackType "human"]

1. d4 Nf6 2. c4 c5 3. d5 e6 4. Nc3 exd5 5. cxd5 d6 6. e4 g6 7. Nf3 Bg7 8.
Bb5+ {(f1b5 b8d7 c1f4 a7a6 b5d3 d8b6 a1b1 e8g8 e1g1 f8d8) +0.50/10 7} Bd7
{(c8d7 d1d3 d7b5 d3b5 d8d7 b5e2 e8g8 e1g1) -0.70/6 8} 9. Bd3 {(b5d3 e8g8
e1g1 b8a6 c1f4 d8b6 a1b1 a8e8 a2a3 a6c7 b2b4) +0.66/11 7} Ng4 {(f6g4 e1g1
g7c3 b2c3 e8g8 c1f4) -0.75/6 8} 10. Qb3 {(d1b3 d8c7 e1g1 e8g8 c1g5 b8a6
c3b5 c7b6 d3c4 g4e5) +0.80/10 8} Qc7 {(d8c7 e1g1 e8g8 c1g5 g4e5 d3e2)
-0.86/6 8} 11. Nb5 {(c3b5 d7b5 d3b5 e8f8 e1g1 b8d7 c1f4 a7a6 b5e2 g4f6
b3c4) +0.82/11 8} Bxb5 {(d7b5 b3b5 b8d7 c1f4 a7a6 b5b3 g4e5 f3e5 d7e5)
-0.53/7 8} 12. Bxb5+ {(d3b5 e8f8 e1g1 b8d7 c1f4 g4e5 b5e2 a7a6 a2a3 f7f5
f3d2) +0.76/11 8} Nd7 {(b8d7 e1g1 e8g8 b5d7 c7d7 c1f4 g7d4) -0.54/6 8} 13.
O-O {(e1g1 a7a6 b5e2 e8g8 b3c2 g4e5 c1f4 a8d8 a2a3 b7b5) +0.50/10 8} O-O
{(e8g8 c1f4 f8e8 b5d7 c7d7 f1e1 g4e5) -0.44/7 8} 14. Bf4 {(c1f4 a7a6 b5e2
g4f6 e2d3 a8d8 a2a3 d7e5 d3c2) +0.52/9 8} Rae8 {(a8e8 b5d3 d7e5 f4e5 g4e5
d3e2) -0.39/6 8} 15. Rae1 {(a1e1 a7a6 b5d7 c7d7 b3b6 g4e5 f4e5 g7e5 f3e5)
+0.50/9 8} Re7 {(e8e7 b5d7 c7d7 f4g5 f7f6 g5f4) -0.47/6 7} 16. a4 {(a2a4
f8e8 b3c2 g4e5 b5e2 d7f6 c2b1 a7a6 b2b4) +0.54/9 7} Rfe8 {(f8e8 b3c2 g4e5
f4e5 g7e5 b2b4) -0.35/6 8} 17. Bd3 {(b5d3 g4e5 d3e2 d7f6 f3d2 a7a6 b3c2
h7h6 b2b3) +0.40/9 7} Be5 {(g7e5 f3e5 g4e5 d3b5 f7f6 e1c1) -0.26/6 8} 18.
Nxe5 {(f3e5 g4e5 d3e2 a7a6 f4g5 f7f6 g5f4 f6f5 f4g5 c5c4) +0.88/10 8} Ngxe5
{(g4e5 d3b5 f7f5 b5d7 c7d7 e4f5 g6f5 f4e5 e7e5) -0.29/7 8} 19. Bc2 {(d3c2
a7a6 e1e2 c7b8 f1d1 d7f6 a4a5 e7d7 d1d2 e8e7 b3e3) +0.88/11 8} Qa5 {(c7a5
b3a3 e5c4 a3a2 d7e5 b2b3 c4d2) -0.25/7 8} 20. Re2 {(e1e2 b7b6 f1d1 a7a6
f4g5 f7f6 g5f4 b6b5 a4b5) +0.94/9 8} f6 {(f7f6 f1c1 c5c4 b3e3 d7c5 f4e5
e7e5) -0.22/7 7} 21. Rfe1 {(f1e1 g6g5 b3g3 a7a6 b2b3 b7b5 a4b5 a6b5 e1d1)
+0.96/9 8} c4 {(c5c4 b3a3 d7c5 e1d1 a5a6 d1c1 c4c3) -0.15/7 8} 22. Qa3
{(b3a3 d7c5 b2b4 c4b3 c2b3 c5b3 a3b3 f6f5) +0.96/8 8} Qa6 {(a5a6 a3b4 d7c5
f4e5 e7e5 b2b3) -0.14/6 8} 23. b4 {(b2b4 b7b5 a4a5 g6g5 f4d2 a6c8 f2f4 c4c3
d2c3) +1.06/9 8} b5 {(b7b5 e1a1 b5a4 c2a4 e8b8 a4d7 a6a3 a1a3 e7d7) +0.43/7
7} 24. a5 {(a4a5 g6g5 f4d2 a6c8 a3e3 a7a6 d2c3 g5g4 c3d4) +1.04/9 7} Qb7
{(a6b7 f4e5 d7e5 f2f4 e5g4 e1c1) +0.45/6 8} 25. a6 {(a5a6 b7c8 e2e3 g6g5
e3g3 c8c7 f4d2 g5g4 d2f4) +1.08/9 8} Qb6 {(b7b6 e1c1 e5f7 f4e3 b6c7 c2d3
d7e5) +0.42/7 8} 26. Be3 {(f4e3 b6b8 e3d4 g6g5 a3e3 e5g4 e3f3 g4e5 f3f5)
+1.26/9 8} Qc7 {(b6c7 e1c1 e5f7 f2f4 d7b6 a3c3 c7b8) +0.39/7 8} 27. f4
{(f2f4 e5f7 e3d4 g6g5 f4f5 f7e5 a3e3 g5g4 e1a1 d7b6 a1e1) +1.30/11 7} Ng4
{(e5g4 e3d4 f6f5 a3b2 d7b6 c2b1 g4h6) +0.43/7 8} 28. Bd4 {(e3d4 f6f5 a3c3
c7b8 h2h3 g4h6 c3f3 b8c8 f3a3 c8c7) +1.38/10 8} f5 {(f6f5 a3a1 c7b8 c2b1
d7b6 a1b2 g4h6) +0.43/7 8} 29. Qc3 {(a3c3 c7b8 h2h3 g4h6 c3f3 b8c7 h3h4
f5e4 c2e4 h6f7) +1.36/10 8} Kf7 {(g8f7 c3f3 g4f6 e4f5 e7e2 e1e2 e8e2 f3e2
f6d5 f5g6 h7g6) +0.48/7 8} 30. Qf3 {(c3f3 f7f8 e4f5 e7e2 e1e2 e8e2 f3e2
g6f5 c2f5 g4f6 e2e6) +2.46/11 7} Ngf6 {(g4f6 e4f5 e7e2 e1e2 e8e2 f3e2 f6d5
f5g6 h7g6) +0.48/7 8} 31. exf5 {(e4f5 e7e2 e1e2 g6f5 e2e8 f7e8 c2f5 e8f8
f5e6 h7h6 f4f5) +2.46/11 8} Rxe2 {(e7e2 e1e2 e8e2 f3e2 d7f8 f5g6 h7g6 e2f3
f7e7) -0.81/8 7} 32. Rxe2 {(e1e2 g6f5 e2e8 f7e8 c2f5 e8f8 f5e6 h7h6 f4f5)
+2.46/9 8} Rxe2 {(e8e2 f3e2 d7f8 e2f3 c7c8 d4a7 c8a6 a7d4) -0.66/8 8} 33.
Qxe2 {(f3e2 d7f8 f5g6 h7g6 f4f5 g6f5 e2f2 c7c8 c2f5 c8a6 f5g4 f8h7)
+2.84/12 7} Nf8 {(d7f8 f5g6 h7g6 f4f5 g6f5 e2f2 f7e7 f2h4 f8d7 c2f5)
-0.90/8 7} 34. fxg6+ {(f5g6 h7g6 e2f3 c7e7 d4c3 f7g8 f4f5 g6g5 c3f6 e7f6
c2e4) +2.66/11 8} hxg6 {(h7g6 f4f5 g6f5 e2f2 f7e7 f2f5 f8d7 g2g4) -1.11/7
8} 35. Qf3 {(e2f3 c7e7 d4c3 f7g8 f4f5 g6g5 c3f6 e7f6 c2e4) +2.66/9 8} Qc8
{(c7c8 f4f5 g6f5 d4a7 c8a6 a7d4 f7e7 c2f5) -0.84/7 8} 36. Bxa7 {(d4a7 c8a6
a7d4 a6a2 f3e2 a2a3 d4f6 f7f6 e2d2 c4c3 d2d4) +2.62/11 7} Qxa6 {(c8a6 a7d4
a6a2 d4f6 a2c2 f6c3 f7e7 g1f1) -0.28/8 8} 37. Bd4 {(a7d4 a6a2 f3f2 a2a3
d4f6 f7f6 f2d4 f6f7 d4d2 f7g8 c2e4) +2.48/11 8} Qa2 {(a6a2 d4f6 a2c2 f6c3
f8d7 g1f1 f7e7 f1e1) -0.28/8 8} 38. Qf2 {(f3f2 a2a3 d4f6 f7f6 f2e1 f6g7
c2e4 c4c3 e1b1 a3b2 b1b2) +2.20/11 7} Qa3 {(a2a3 d4f6 a3c1 f2f1 c1c2 f6e5
d6e5 f4e5 f7e7 d5d6) +0.03/8 8} 39. Bxf6 {(d4f6 a3c1 f2f1 c1c2 f6h4 c2c3
f4f5 c3d4 f1f2 d4d1 f2f1 d1f1) +1.60/12 8} Qc1+ {(a3c1 f2f1 c1c2 f6e5 f7e7
f1e1 d6e5 f4e5 c4c3) +0.42/8 8} 40. Qf1 {(f2f1 c1c2 f6h4 c2c3 f4f5 c3d4
f1f2 d4a1 f2f1 a1f1 g1f1 g6f5) +1.60/12 7} Qxc2 {(c1c2 f6e5 f7e7 f1e1 d6e5
f4e5 c4c3 d5d6) +0.26/8 7} 41. Qa1 {(f1a1 c2d3 f6h8 d3e3 g1h1 e3f4 a1g7
f7e8 g7d4 f4f1 d4g1) +1.50/11 8} Qd2 {(c2d2 f6g7 d2e3 g1h1 e3f4 a1a7 f7g8
g7b2 f4g5 a7b7) +0.73/8 7} 42. Bd4 {(f6d4 d2b4 a1a7 f7e8 a7a8 e8e7 a8b7
f8d7 d4f6 e7f6 b7d7) +1.36/11 8} Qxb4 {(d2b4 g2g4 b4d2 f4f5 g6f5 g4f5 b5b4)
+0.96/7 7} 43. Qa7+ {(a1a7 f7e8 a7a8 e8e7 a8a1 f8d7 g2g4 d7c5 f4f5 g6f5
g4f5) +1.46/11 8} Ke8 {(f7e8 g1f1 b4b1 f1f2 b1c2 f2f3 c2d3 f3f2 b5b4 a7b8
e8e7 b8c7) +1.31/7 8} 44. Qa8+ {(a7a8 e8e7 a8a1 f8d7 a1d1 e7f8 d1a1 f8g8
a1a8 d7f8) +1.34/10 8} Ke7 {(e8e7 d4f6 e7f7 f6h4 b4b1 g1f2 b1c2 f2f3 c2d3
f3f2 b5b4 h4g3) +1.31/7 8} 45. Qa1 {(a8a1 b4d2 g2g3 f8d7 h2h3 d2d3 a1e1
e7f8 e1c3 f8g8) +1.08/10 8} Qd2 {(b4d2 g2g3 f8d7 h2h3 b5b4 g1f1 b4b3 a1e1)
+1.70/8 8} 46. Bf6+ {(d4f6 e7f7 f6d4 d2f4 a1a7 f7e8 d4b2 f8d7 a7a5 f4e3
g1f1 e3d3) +0.62/12 7} Ke8 {(e7e8 f6d4 f8d7 g2g3 b5b4 h2h3 c4c3 a1a8)
+1.63/7 7} 47. Bg7 {(f6g7 d2d5 g7f8 e8f8 a1f6 d5f7 f6d6 f8g7 d6c5 g7g8
c5b5) +0.38/11 8} Nd7 {(f8d7 g7h6 c4c3 g1f1 b5b4 a1a8 e8f7 a8b7 d2c1)
+1.76/8 8} 48. Bh6 {(g7h6 c4c3 g1f1 d7b6 a1b1 b6d5 b1b5 e8d8 h6g5 d8c8)
+0.58/10 8} c3 {(c4c3 h2h4 d7b6 h4h5 d2d4 g1h1 g6h5 a1e1) +2.22/7 8} 49. h3
{(h2h3 d7b6 a1a7 d2d1 g1h2 b6d5 a7b8 e8f7 b8b5 f7g8 h6g5) +0.08/11 8} Nb6
{(d7b6 a1a7 b6d5 a7b8 e8e7 h6f8 e7f6 b8d6 f6f5 g2g3) +2.01/7 8} 50. Qa7
{(a1a7 d2d1 g1h2 b6d5 a7b8 e8d7 b8b5 d7c7 b5c4 c7b7 h6g5 c3c2) +0.08/12 8}
Qd1+ {(d2d1 g1h2 b6d5 a7b8 e8f7 b8f8 f7e6 f8e8 e6f5 g2g4 f5f6 e8b5) +1.63/7
8} 51. Kh2 {(g1h2 b6d5 a7b8 e8d7 b8b5 d7c7 h6g5 c3c2 b5c4 c7b7 c4b3 b7c8
b3c4) -0.02/13 8} Nxd5 {(b6d5 a7b8 e8d7 b8b5 d7c7 b5c4 c7d8 h6g5 d8d7 c4b5
d7e6 b5e8 e6f5 h2g3) +1.47/8 8} 52. Qb8+ {(a7b8 e8d7 b8b5 d7c7 b5c4 c7b7
h6g7 c3c2 g7b2 b7a7 b2d4 a7b8) -0.10/12 8} Kd7 {(e8d7 b8b5 d7c7 b5c4 c7d8
h6g5 d8d7 c4b5 d7c7 b5c4 c7b6 g5d8 b6b7 h2g3) +1.27/7 7} 53. Qxb5+ {(b8b5
d7c7 h6g5 c3c2 b5c4 c7b8 c4b5 b8c8 b5c4 c8b7 c4b3 b7c8 b3c4) -0.02/13 8}
Kc7 {(d7c7 b5c4 c7d8 c4c6 c3c2 c6d6 d8c8 d6e6 c8c7 f4f5 g6f5 e6f5) +0.02/7
8} 54. Qc4+ {(b5c4 c7b7 h6g7 c3c2 g7b2 d1d2 c4b5 b7c7 b5c4 c7b8 c4b5 b8c8)
-0.04/12 8} Kd8 {(c7d8 c4c6 c3c2 c6d6 d8c8 d6e6 c8c7 f4f5 g6f5 e6f5)
+0.02/7 7} 55. Qc6 {(c4c6 c3c2 c6d6 d8c8 d6e6 c8c7 f4f5 g6f5 e6f5 c7c6 f5e6
c6c5 h6f8) +0.66/13 8} c2 {(c3c2 f4f5 d5e7 c6b6 d8d7 b6b7 d7d8 b7b8 d8d7
b8b7) 0.00/8 7} 56. f5 {(f4f5 g6f5 c6d6 d8c8 d6e6 c8c7 e6f5 c7c6 f5e6 c6c5
h6g5 d1b1 e6c8 c5d4) +0.84/14 8} Ne7 {(g6f5 h6g5 d5e7 c6a8 d8c7 a8a7 c7c6
a7e7 c2c1q g5c1 d1c1) -0.07/7 8} 57. Qb6+ {(c6b6 d8c8 f5f6 c2c1q h6c1 d1c1
f6e7 c1f4 h2h1 f4c1 h1h2 c8d7 b6b7) +2.32/13 8} Ke8 {(d8e8 f5g6 e7g6 b6b5
e8d8 b5a5 d8e8 a5b5) 0.00/7 7} 58. Qb8+ {(b6b8 e8d7 b8b7 d7e8 f5f6 d1e2
b7e7 e2e7 f6e7 e8e7 h6f4 d6d5 h2g1) +2.72/13 8} Kd7 {(e8d7 b8b7 d7d8 f5f6
d1e2 f6e7 e2e7 b7e7 d8e7 h6g5 e7e6 h2g3) -0.34/7 8} 59. Qb7+ {(b8b7 d7e8
f5f6 d1e1 b7e7 e1e7 f6e7 e8e7 h2g1 d6d5 g2g4 d5d4 h6f4) +2.60/13 7} Ke8
{(d7e8 f5f6 d1e2 f6e7 e2e5 g2g3 e5e2 h2g1 e2e7 b7d5) -0.34/7 8} 60. f6
{(f5f6 d1e2 f6e7 e2e5 g2g3 e5e2 h2g1 e2d1 g1g2 c2c1q h6c1 d1c1) +2.88/12 8}
Qe2 {(d1e2 b7a8 e8d7 f6f7 e2e5 g2g3 e5e2 a8g2 e2g2 h2g2 g6g5 h6g5) -1.41/7
8} 61. fxe7 {(f6e7 e2e5 g2g3 e5c5 h6c1 c5f2 b7g2 f2g2 h2g2 e8e7 c1f4)
+2.78/11 8} Qe5+ {(e2e5 g2g3 e5e2 b7g2 e2g2 h2g2 e8e7 h6g5 e7e6 g2f3 d6d5
g5e3) -0.21/8 8} 62. g3 {(g2g3 e5c5 h6c1 c5f2 b7g2 f2g2 h2g2 e8e7 c1f4 d6d5
g2g1 e7f8) +2.38/12 8} Qe2+ {(e5e2 b7g2 e2g2 h2g2 e8e7 h6d2 d6d5 g2f3 d5d4
f3e4) -0.22/8 8} 63. Qg2 {(b7g2 e2g2 h2g2 e8e7 g2f1 e7f7 h6f4 d6d5 f1e2
f7f6 e2d2 g6g5) +2.46/12 8} Qxg2+ {(e2g2 h2g2 e8e7 g2f3 e7f7 f3e4 f7g8 e4d5
g8h7) -1.04/8 7} 64. Kxg2 {(h2g2 e8e7 g2f1 d6d5 f1e2 d5d4 e2d2 e7d6 d2c2
d6d5 c2b1 d5e4 g3g4 d4d3 h6g5) +3.58/15 8} Kxe7 {(e8e7 g2f3 e7f6 f3e2 g6g5
e2d2 g5g4 h3g4 f6g6) -1.99/9 8} 65. Kf2 {(g2f2 d6d5 f2e2 d5d4 e2d2 e7d6
d2c2 d6d5 c2b1 d5e4 g3g4 d4d3 h6g5) +3.58/13 8} Kf6 {(e7f6 h6d2 d6d5 f2e3
f6e5 e3d3 c2c1q d2c1 d5d4) -2.19/9 7} 66. Ke1 {(f2e1 f6e5 h6f4 e5e4 e1d2
d6d5 d2c2 d5d4 c2b2 e4d5 g3g4 d4d3 b2b1 d5e4) +3.60/14 8} d5 {(d6d5 e1d2
c2c1r d2c1 f6e5 c1d2 e5e4 h6g5 d5d4 g5f4) -2.34/10 7} 67. Bc1 {(h6c1 f6f5
e1d2 f5e5 d2c2 e5e4 c2b1 d5d4 c1f4 d4d3 f4g5 e4d4 g5f6 d4e4 g3g4) +3.58/15
8} Ke5 {(f6e5 e1d2 e5e4 d2c2 e4d4 c1f4 d4e4 c2b3 d5d4 b3c4) -2.44/10 8} 68.
Kd2 {(e1d2 e5d4 c1b2 d4e4 d2c2 d5d4 c2b1 d4d3 g3g4 e4e3 b2c3 e3f4 c3e1 f4e4
e1c3) +3.58/15 8} Kd4 {(e5d4 d2c2 d4e4 g3g4 e4d4 h3h4 d4e4 c2d2 d5d4 h4h5
g6h5 g4h5) -2.51/11 8} 69. Kxc2 {(d2c2 d4e4 c1f4 e4d4 c2b1 d4c4 g3g4 d5d4
h3h4 d4d3 h4h5 g6h5 g4h5 c4d4 h5h6) +3.72/15 8} Ke4 {(d4e4 c2c3 e4f3 c1g5
f3e4 g5f4 e4f5 c3d4 f5e6 f4e5) -2.93/10 8} 70. Bf4 {(c1f4 e4d4 c2d2 d4c4
f4e5 c4b4 g3g4 b4b3 d2d3 b3a2 g4g5 a2b1 d3d2 b1a2 d2c1) +4.34/15 7} Kd4
{(e4d4 h3h4 d4c4 f4e5 c4c5 c2d3 c5b6 d3d4 b6c6 g3g4) -3.06/10 8} 71. h4
{(h3h4 d4c4 f4e5 d5d4 e5f6 d4d3 c2d2 c4d5 d2d3 d5e6 f6d4 e6f5 d3c4 f5e4
g3g4) +4.98/15 8} Ke4 {(d4e4 c2c3 d5d4 c3d2 d4d3 d2c3 d3d2 c3d2 e4d4 g3g4
d4e4) -4.09/10 7} 72. Kc3 {(c2c3 d5d4 c3c4 d4d3 c4c3 e4f5 c3d3 f5e6 d3d4
e6e7 d4e4 e7f7 g3g4 f7f8 h4h5 g6h5 g4h5) +5.10/17 8} d4+ {(d5d4 c3c2 e4d5
c2d3 d5c5 f4e5 c5d5 e5d4 d5d6 d3e4 d6e6) -4.34/10 8} 73. Kc4 {(c3c4 d4d3
c4c3 d3d2 c3d2 e4d4 d2c1 d4e4 f4g5 e4d4 c1b1 d4e4 g5f4 e4d5 g3g4 d5e4 f4g5
e4d4) +5.16/18 8} d3 {(d4d3 c4c3 d3d2 c3d2 e4d4 g3g4 d4e4 f4e3 e4e5 h4h5
g6h5 g4h5) -4.48/11 8} 74. Kc3 {(c4c3 d3d2 f4d2 e4d5 c3b2 d5d4 d2f4 d4e4
b2b1 e4f3 f4e5 f3e4 e5f4 e4f3 f4e5 f3e3 g3g4 e3e4 e5b2) +5.16/19 8} d2
{(d3d2 c3d2 e4d4 f4e3 d4e5 g3g4 e5d6 h4h5 g6h5 g4h5 d6e6 h5h6 e6d5)
-4.69/12 7} 75. Kxd2 {(c3d2 e4d4 d2c1 d4e4 f4g5 e4d3 c1b1 d3e4 g5f4 e4f3
f4e5 f3e4 e5f4 e4d5 g3g4 d5e4 f4g5) +5.16/17 7} Kf5 {(e4f5 d2d3 f5e6 d3e4
e6e7 g3g4 e7d7 f4e5 d7e6 h4h5 g6h5 g4h5) -4.83/11 8} 76. Kd3 {(d2d3 f5e6
d3e4 e6d7 g3g4 d7e7 f4e5 e7e6 e5b2 e6d7 e4f4 d7e6 f4e4 e6d7 h4h5 g6h5 g4h5)
+5.38/17 7} Ke6 {(f5e6 d3e4 e6d7 g3g4 d7e7 h4h5 g6h5 g4h5 e7e6 h5h6)
-4.99/10 8} 77. Kd4 {(d3d4 e6f7 d4e4 f7g8 g3g4 g8f8 e4d5 f8e7 d5d4 e7f8
d4e5 f8g7 e5e6 g7g8 e6f6 g8h7) +5.38/16 8} Kf6 {(e6f6 g3g4 f6e6 f4e5 e6d7
h4h5 g6h5 g4h5 d7e6 h5h6) -5.04/10 8} 78. g4 {(g3g4 f6f7 d4e4 f7f8 e4d5
f8g8 d5e5 g8f7 f4h6 f7e7 e5f4 e7e6 f4e4 e6d7 h6f4) +5.38/15 7} Ke6 {(f6e6
h4h5 g6h5 g4h5 e6f6 h5h6 f6f7 f4e5 f7g6 e5f4) -5.19/10 7} 79. Ke4 {(d4e4
e6e7 e4d5 e7f7 f4e5 f7e7 e5d4 e7f8 d5e5 f8g7 e5e6 g7g8 e6f6 g8h7) +5.38/14
8} Ke7 {(e6e7 h4h5 g6h5 g4h5 e7f7 h5h6 f7g6 e4e5 g6h5) -5.19/9 8} 80. Ke5
{(e4e5 e7f7 g4g5 f7e7 f4e3 e7f7 e3c5 f7g7 e5e6 g7g8 e6f6 g8h8 f6g6 h8g8
g6f5) +5.88/15 8} Kf7 {(e7f7 f4g5 f7e8 e5e6 e8f8 e6f6 f8e8 g5f4 g6g5 f6g5
e8e7) -6.31/11 7} 81. Bg5 {(f4g5 f7g7 g5f4 g7f7 f4g5 f7g7 e5e6 g7g8 e6f6
g8h7 g5f4 h7g8 f6g6 g8f8 g6h7 f8e7) +6.02/16 8} Kg7 {(f7g7 e5e6 g7f8 g5f4
f8e8 g4g5 e8f8 e6f6 f8e8 f4e5 e8d7 f6g6) -6.40/11 8} 82. Ke6 {(e5e6 g7h8
g5f6 h8g8 e6e7 g6g5 f6g5 g8g7 h4h5 g7h8 g5f6 h8g8 f6e5 g8h7 e7f7 h7h6)
+6.36/16 8} Kg8 {(g7g8 e6f6 g8h7 f6f7 h7h8 f7g6 h8g8 g5e7 g8h8 e7f6 h8g8
g6f5) -6.71/11 7} 83. Kf6 {(e6f6 g8h7 g5f4 h7g8 f6g6 g8f8 h4h5 f8e7 h5h6
e7e6 h6h7 e6d5 h7h8q d5e4 h8e5 e4d3 g6f5 d3c4) +14.28/18 8} Kh7 {(g8h7 f6f7
h7h8 h4h5 g6h5 g5f6 h8h7 g4g5 h5h4 g5g6 h7h6 g6g7 h4h3 g7g8q) -10.31/11 8}
84. Bf4 {(g5f4 h7h8 f6g6 h8g8 h4h5 g8f8 h5h6 f8e8 f4g3 e8e7 h6h7 e7d7 h7h8q
d7e6 g3e5 e6d5 g6f5 d5c5 f5e4) +14.36/19 7} Kg8 {(h7g8 f6g6 g8f8 h4h5 f8e7
h5h6 e7e6 h6h7 e6d5 g6f5 d5d4 h7h8q) -13.39/11 7} 85. Kxg6 {(f6g6 g8f8 h4h5
f8e7 h5h6 e7e8 g6f6 e8d7 h6h7 d7c6 h7h8q c6d5 g4g5 d5e4 f4e5 e4d5 g5g6 d5e4
g6g7) +14.70/19 7} Kf8 {(g8f8 h4h5 f8e7 h5h6 e7e6 h6h7 e6d5 g6f5 d5d4 h7h8q
d4d5 g4g5) -13.61/11 8} 86. Be5 {(f4e5 f8e7 h4h5 e7e6 e5a1 e6d5 a1f6 d5e4
h5h6 e4f4 g4g5 f4e3 h6h7 e3f4 h7h8q f4e4 f6d4) +14.00/17 8} Ke7 {(f8e7 e5f6
e7d7 g4g5 d7c8 h4h5 c8b7 h5h6 b7c6 h6h7 c6d5 h7h8q) -13.41/11 8 Arena
Adjudication} 1-0
[/pgn]

Code: Select all

s32 wkPST[64] = {
   10,  40,  30, -40, -30,  30,  40,  10,
  -10,  20, -10,   0,   0, -10,  20, -10,
  -20, -20, -20, -20, -20, -20, -20, -20,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  20,  20,  20,   0, -10,
  -10,   0,   0,   0,   0,   0,   0, -10,
  -10, -10, -10, -10, -10, -10, -10, -10
};

[lithander]

Code: Select all

s32 wkPST[64] = {
   10,  40,  30, -40, -30,  30,  40,  10,
  -10,  20, -10,   0,   0, -10,  20, -10,
  -20, -20, -20, -20, -20, -20, -20, -20,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  20,  20,  20,   0, -10,
  -10,   0,   0,   0,   0,   0,   0, -10,
  -10, -10, -10, -10, -10, -10, -10, -10
};

I replaced the KING table with this one and ran some tests again and the tables are definitely an improvement, but not a massive one. (~20 ELO) By now I'm almost convinced that the approach of using static PSTs as the only term in the evaluation is just too simplistic.

Code: Select all

Score of MinimalChess 0.3 vs MinimalChess 0.3 Carnivor: 334 - 364 - 302  [0.485] 1000
Elo difference: -10.4 +/- 18.0, LOS: 12.8 %, DrawRatio: 30.2 %

Score of Rustic vs MinimalChess 0.3 Carnivor: 571 - 255 - 174  [0.658] 1000
Elo difference: 113.7 +/- 20.4, LOS: 100.0 %, DrawRatio: 17.4 %

Have you tested against TSCP1.81? If so can you post the results for the various PSTs.

I have run tests against TSCP1.81 but not with all PST sets. Only the default one. That you can put PSTs from other programs in a text file and MMC will play by them is just a gimmick feature. It takes too long if I repeat all tests 4x or 5x times so that I know exactly how a certain set of PSTs performs. To be honest I wasn't even aware of the fundamental problems my current evaluation has in endgame situations until Marcel pointed that out...

One of the reasons your PST above can't convert wins is because the king gets stuck on the 7th rank. There is not enough incentive to leave it, so you can't mate with K+R vs. K because you can't see 15-20 moves ahead (yet). What you need, is a table where the center is 0, and all of the other squares are negative. Thus, the king wants to be in the center if he is not there already. Then, the other king needs to be driven to the edge.

...and after that I reviewed a few PGNs and setup a few cases where I gave MinimalChess a Queen and King vs a lone King, or a Rook or two Bishops and MMC was just failing to win the game regardless of what PSTs I used. Even the sunfish ones didn't convert.

I tested the same setup with weaker engines and some of them played same situations flawlessly. That's very disappointing but I'm not sure how to address it, yet. The evaluation part is harder to do for me because it requires chess knowledge and not just programming experience. While I understand how the PST values influence the way the engine plays I don't know why the values are what they are in the PSTs I've used so far. And this caused the above mentioned problems without me even being aware of them. The consequence for me is that I need to "own" that part of my engine too, even if at first it means a regression in playing strength.

Just my opinion but 20s as high values on the 7th rank is too close to the eighth rank if the engine has any depth at all. However, it seems like your test are at very fast time controls so maybe, idk, the depth your engine reaches is not enough to see that if the king is in the center that it can go to the 7th to help a pawn promote. I think what I'm trying to say is that if the engine searches deep enough the high scores should be in the center so it can head anywhere on the board as needed. Why do the PSTs have to be 100% static? Can't we just work from copies and modify the copies according to the game phase? Or do like I did in RomiChess just compute the PSTs at the root. So just copy the static tables at the root and then adjust the values with mobility for a minimal engine and anything that can be thought of for a more advanced engine.

The time controls are 3 seconds + 500ms per move. The PSTs don't "have" to be static but that was the first thing I've tried. Before that I was only counting material, when I tried PSTs the playing strength improved massively and I thought that meant the evaluation was doing it's job. I have to admit that I don't really understand everything you said above. The only conclusion I'm ready to make at this point is that I need to study this topic more. Like actually reading books on chess.

If you have any tips regarding where to start reading let me know. (You mentioned mobility for example)

[mvanthoor]

Have you tested against TSCP1.81? If so can you post the results for the various PSTs. I recommend TSCP because it does have a standard and decent evaluation function. It might result in better test data because it would not be as likely to fall into the trap a > b > c > a when testing PSTs because all static PSTs will be strong in some positions and weak in others.

No it doesn't... TSCP is fun to watch, but it's a one-trick pony, whose trademark is destroying pawn structures and creating passed pawns. Even Alpha 2, which will probably be rated higher than TSCP after testing, will also probably score a lower rating against TSCP than expected, because it doesn't have any pawn-related evaluations yet. If you add passed pawns (and how dangerous they are) and try to prevent double/triple pawns, it'll gain you somewhere between 100-125 Elo, and suddenly your results against specifically TSCP will skyrocket.

[mvanthoor]

...and after that I reviewed a few PGNs and setup a few cases where I gave MinimalChess a Queen and King vs a lone King, or a Rook or two Bishops and MMC was just failing to win the game regardless of what PSTs I used. Even the sunfish ones didn't convert.

I tested the same setup with weaker engines and some of them played same situations flawlessly. That's very disappointing but I'm not sure how to address it, yet. The evaluation part is harder to do for me because it requires chess knowledge and not just programming experience. While I understand how the PST values influence the way the engine plays I don't know why the values are what they are in the PSTs I've used so far. And this caused the above mentioned problems without me even being aware of them. The consequence for me is that I need to "own" that part of my engine too, even if at first it means a regression in playing strength.

I experienced the same, but the other way around.

I didn't have a lot of experience in chess programming a year ago, and when I got to the point where MMC is now about 6 months ago, I had the same issue as you. For me, as a chess player, mating with K+Q vs. K and K+R vs. K is so completely trivial, that I couldn't at first understand why my engine would be that stupid.

The reason why it doesn't work, is because the engine can't see 10-20 moves ahead; it can't see the mate from the position it is in. It doesn't know, however, that the lone king must be at the edge to mate it. If you just use your midgame/opening PST, your king will be stuck in the corner it castled to, and the engine will keep playing around with the queen in the middle of the board.

After HGM pointed out to me that I need an endgame PST with massive penalties that become bigger with each step, just to teach the engine that a king on the edge in the endgame is NOT good for the lone king, it was easy.

The only two changes I made were:
- Create the king endgame PST Rustic now has
- Switch to that PST as soon as one of the sides has an evaluation of less than 1 pawn (so only the king is on the board; no squares in any PST has a value larger than or smaller than +99 or -99).

It makes the engine play a bit beginner-like sometimes. I can see a mate coming, but the engine can't, because it has no other evaluation except material and PST's. In such a case the engine will whittle the opponent down to a bare king, piece by piece (because that is what it knows how to do), and then the king endgame table kicks in. The strong king will race to the center, and from there, will proceed to help mate the other king.

As said, it's a beginner-like way of playing such endgames, but that doesn't matter. If you use Rustic's endgame PST and then switch to that if the evaluation is less than 100 cp for one player, the engine should convert K+Q vs. King and K+R vs. King without issue. Mine did; it gained about 60 Elo by not botching games.

And you're right: to understand the PST's I wrote, you have to have some chess knowledge. Quite a bit of positional knowledge is written into them; there's even some opening theory in there. Sunfish and other tables? You can't understand them. I can't either. They contain knowledge of the evaluation function of whatever helped tune them.

I could explain the PST values in general terms; it would also make a good piece for the tutorial site I'm writing

Also: don't put too much effort in your evaluation function. First:
- Get move sorting in order (pun intended). Have PV Move, Killer, History ordering.
- Add a transposition table and add PV Move ordering

Then your basic engine will be done, and as you can see with Rustic, it will already play close to 1800 Elo. (Rustic doesn't even have PV Move, Killer, and History ordering yet... I wonder how much they will gain.)

At this point, I'll be looking into Texel Tuning to make the basic engine as strong as possible, and so I don't have to tinker with the PST's any longer by myself. If you have some good PST's, you can skip this step for later. I don't, because the next steps are:

- Tapered evaluation, which can then be Texel tuned. Big Elo gain by both the tapering, and the tuning, without any chess knowledge needed.

- Add Aspiration Windows to iterative deepening / alpha-beta (speed-up)
- Add Principal Variation Search to alpha-beta (speed-up)
- Add null move (massive speed-up)
- Add as many other pruning options as you can understand; one by one. Futility pruning, mate distance pruning, WhatEverPruning... I must admit, I still have to study many of those.
- Keep re-tuning the tapered evaluation as you improve the search function.

This is the path that was taken by the PESTO engine, and it reached a massive strength doing so. It requires studying algorithms, and it requires programming experience and lots of testing, but no chess knowledge. So you can take your time and actually study some chess. Then at the end you can start implementing the easier evaluation terms.

The trick is to go slow, and do only ONE function at a time, and not try to get carried away in an urge to get your engine as high up in the rating list as possible. Understand the code you write and why you write it; the evaluation, which requires only minimal programming knowledge and a lot of chess knowledge, can wait... precisely because it can be tuned.

[Mike Sherwin]

Code: Select all

s32 wkPST[64] = {
   10,  40,  30, -40, -30,  30,  40,  10,
  -10,  20, -10,   0,   0, -10,  20, -10,
  -20, -20, -20, -20, -20, -20, -20, -20,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  30,  30,  20,   0, -10,
  -10,   0,  20,  20,  20,  20,   0, -10,
  -10,   0,   0,   0,   0,   0,   0, -10,
  -10, -10, -10, -10, -10, -10, -10, -10
};

I replaced the KING table with this one and ran some tests again and the tables are definitely an improvement, but not a massive one. (~20 ELO) By now I'm almost convinced that the approach of using static PSTs as the only term in the evaluation is just too simplistic.

Code: Select all

Score of MinimalChess 0.3 vs MinimalChess 0.3 Carnivor: 334 - 364 - 302  [0.485] 1000
Elo difference: -10.4 +/- 18.0, LOS: 12.8 %, DrawRatio: 30.2 %

Score of Rustic vs MinimalChess 0.3 Carnivor: 571 - 255 - 174  [0.658] 1000
Elo difference: 113.7 +/- 20.4, LOS: 100.0 %, DrawRatio: 17.4 %

Have you tested against TSCP1.81? If so can you post the results for the various PSTs.

I have run tests against TSCP1.81 but not with all PST sets. Only the default one. That you can put PSTs from other programs in a text file and MMC will play by them is just a gimmick feature. It takes too long if I repeat all tests 4x or 5x times so that I know exactly how a certain set of PSTs performs. To be honest I wasn't even aware of the fundamental problems my current evaluation has in endgame situations until Marcel pointed that out...

One of the reasons your PST above can't convert wins is because the king gets stuck on the 7th rank. There is not enough incentive to leave it, so you can't mate with K+R vs. K because you can't see 15-20 moves ahead (yet). What you need, is a table where the center is 0, and all of the other squares are negative. Thus, the king wants to be in the center if he is not there already. Then, the other king needs to be driven to the edge.

...and after that I reviewed a few PGNs and setup a few cases where I gave MinimalChess a Queen and King vs a lone King, or a Rook or two Bishops and MMC was just failing to win the game regardless of what PSTs I used. Even the sunfish ones didn't convert.

I tested the same setup with weaker engines and some of them played same situations flawlessly. That's very disappointing but I'm not sure how to address it, yet. The evaluation part is harder to do for me because it requires chess knowledge and not just programming experience. While I understand how the PST values influence the way the engine plays I don't know why the values are what they are in the PSTs I've used so far. And this caused the above mentioned problems without me even being aware of them. The consequence for me is that I need to "own" that part of my engine too, even if at first it means a regression in playing strength.

Just my opinion but 20s as high values on the 7th rank is too close to the eighth rank if the engine has any depth at all. However, it seems like your test are at very fast time controls so maybe, idk, the depth your engine reaches is not enough to see that if the king is in the center that it can go to the 7th to help a pawn promote. I think what I'm trying to say is that if the engine searches deep enough the high scores should be in the center so it can head anywhere on the board as needed. Why do the PSTs have to be 100% static? Can't we just work from copies and modify the copies according to the game phase? Or do like I did in RomiChess just compute the PSTs at the root. So just copy the static tables at the root and then adjust the values with mobility for a minimal engine and anything that can be thought of for a more advanced engine.

The time controls are 3 seconds + 500ms per move. The PSTs don't "have" to be static but that was the first thing I've tried. Before that I was only counting material, when I tried PSTs the playing strength improved massively and I thought that meant the evaluation was doing it's job. I have to admit that I don't really understand everything you said above. The only conclusion I'm ready to make at this point is that I need to study this topic more. Like actually reading books on chess.

If you have any tips regarding where to start reading let me know. (You mentioned mobility for example)

I would download the latest Crafty and just read all the comments. Bob explains everything and why. It was 17 years ago that I read it and all I remember is that it was very helpful!