Engine Release: Valiant

[Enderjed]

So far on my PC, the fastest I've been able to get 3.2 (and 3.2 "Archer") is 16 Kn/s, and generally takes less time than 30 seconds for it's first move.
It does search at far greater depths and faster when there are less pieces on the board (so far the current record is a depth of 12) [In testing, this has given Valiant a stronger endgame, giving it a unique skill curve over game time due to it's depth being tied to piece count.]
{Versions older than 3.2 and 3.2 archer run much slower due to their time management code being quite lenient}

The reasoning on why the engine is so slow is because:
1. It's compiled for python 3.x, which runs at the speed of an actual python
2. I would compile it in PyPy 3 to make it go [drastically] faster, but I have yet to learn on how to get it to successfully compile in PyPy3
3. Iterative deepening is what consumes the most time, but without it, the engine would be drastically lower in elo

Although I must say that I do thank you (and alvinypeng) for downloading my engine and giving it a shot.
Is there anything about it (strategically and/or positionally) that could use improvement?

Using tapered piece square tables can improve your engine positionally. Also, you can speed up your engine by incrementally updating your piece square table scores every time you make a move rather than computing them from scratch every time you call your evaluate function.

I did try doing that, but it lead to distorted scores, and terrible moves.

[Whiskers]

So far on my PC, the fastest I've been able to get 3.2 (and 3.2 "Archer") is 16 Kn/s, and generally takes less time than 30 seconds for it's first move.
It does search at far greater depths and faster when there are less pieces on the board (so far the current record is a depth of 12) [In testing, this has given Valiant a stronger endgame, giving it a unique skill curve over game time due to it's depth being tied to piece count.]
{Versions older than 3.2 and 3.2 archer run much slower due to their time management code being quite lenient}

The reasoning on why the engine is so slow is because:
1. It's compiled for python 3.x, which runs at the speed of an actual python
2. I would compile it in PyPy 3 to make it go [drastically] faster, but I have yet to learn on how to get it to successfully compile in PyPy3
3. Iterative deepening is what consumes the most time, but without it, the engine would be drastically lower in elo

Although I must say that I do thank you (and alvinypeng) for downloading my engine and giving it a shot.
Is there anything about it (strategically and/or positionally) that could use improvement?

Using tapered piece square tables can improve your engine positionally. Also, you can speed up your engine by incrementally updating your piece square table scores every time you make a move rather than computing them from scratch every time you call your evaluate function.

I did try doing that, but it lead to distorted scores, and terrible moves.

Distorted scores? Are you sure there wasn't a bug in your code? A ((middlegamescore*phase) + (endgamescore*(maxphase-phase))/maxphase should not be very hard to implement and it gave a couple hundred elo for me.

[Enderjed]

Distorted scores? Are you sure there wasn't a bug in your code? A ((middlegamescore*phase) + (endgamescore*(maxphase-phase))/maxphase should not be very hard to implement and it gave a couple hundred elo for me.

It certainly could have been buggy or straight up logically incorrect code, as I am quite the poor programmer.

[Enderjed]

Due to a workload increase, updates to Valiant will be far less constant.
Although do feel free to enter the engine into tournaments, and/or suggest improvements.

[Whiskers]

I got around to fully testing it and I am pretty sure something is wrong with your piece square tables. It keeps running its king up (and getting immediately crushed).
I looked at the eval code to 3.1 (I don't know if these mistakes are also in 3.2, but by its play it seems to be) and found:

1. [rather minor] your queen pst needs to be flipped like all of your other tables
2. [very major!] you are using the king pst for the endgame at all times, including the middlegame! Your search is not very deep, so what kept happening over and over again was Valiant kept running the king towards the center to improve its static evaluation score, and was unable to see that it lost tactically.

An example game which shows it:
[Event "*"] [Site "*"] [Date "2023.03.05"] [Time "23:08:15"] [Round "*"] [White "Willow 2.5"] [Black "Valiant 3.2"] [Result "1-0"] [Termination "mate"] [ECO "E10"] [Opening "Queen's pawn game"] [TimeControl "60+1"] [PlyCount "51"] 1.d4 Nf6 2.c4 e6 3.Nf3 {End of opening} Nc6 {-0.1/4 3620 8036; E10: Queen's pawn game} 4.e3 {+0.2/12 2229 470107} d5 {-0.1/4 5635 8618} 5.Nc3 {+0.2/12 1944 424338} Bb4 {-0.3/3 1839 122} 6.Bd3 {+0.5/13 4353 934391} Bxc3+ {-0.3/4 5821 9973} 7.bxc3 {+0.7/12 2836 625799} Kd7 {-0.5/3 1557 127} 8.Ba3 {+1.3/12 1806 396985} Ne4 {-0.5/3 2328 3687} 9.Qc2 {+1.4/11 2065 421962} f5 {-0.3/3 2315 3225} 10.cxd5 {+1.6/12 2563 541941} exd5 {-0.4/4 4909 5450} 11.c4 {+1.7/13 1789 362101} Ke6 {-0.5/3 1717 1610} 12.cxd5+ {+5.7/14 2302 428119} Kf6 {+4.3/4 3010 5198} 13.dxc6 {+5.9/14 3297 647607} Qd5 {+4.4/4 6986 7268} 14.Bb2 {+5.9/13 3030 606293} bxc6 {+3.6/3 2397 3136} 15.O-O {+6.0/13 4927 938080} Bb7 {+3.3/3 3150 5336} 16.Bc4 {+6.9/13 1636 225497} Qd6 {+3.6/3 2216 3054} 17.d5+ {+7.2/13 1724 313152} Kg6 {+3.7/4 4843 4492} 18.Ne5+ {+10.7/12 1991 394191} Kf6 {+4.5/4 4304 3246} 19.dxc6 {+12.9/13 3484 694331} Bxc6 {+6.9/3 2841 312} 20.Nf7+ {+13.2/14 2393 540059} Kg6 {+9.2/3 731 1509} 21.Nxd6 {+13.6/13 1880 411201} cxd6 {+9.3/3 1350 2968} 22.f3 {+13.6/11 1596 386518} Nc5 {+10.0/3 1568 3195} 23.Qc3 {+15.3/13 1765 389367} Kh5 {+10.4/3 1299 2321} 24.Qxg7 {+999.9/12 1543 276863} Rhc8 {+11.1/3 2030 5085} 25.Qxh7+ {+1000.0/14 2133 435021} Kg5 {+320.0/7 973 3233} 26.h4# {+1000.0/15 1859 358642} 1-0

Fixing that should definitely boost your engine's rating by a few hundred ELO points

[Enderjed]

I got around to fully testing it and I am pretty sure something is wrong with your piece square tables. It keeps running its king up (and getting immediately crushed).
I looked at the eval code to 3.1 (I don't know if these mistakes are also in 3.2, but by its play it seems to be) and found:

1. [rather minor] your queen pst needs to be flipped like all of your other tables
2. [very major!] you are using the king pst for the endgame at all times, including the middlegame! Your search is not very deep, so what kept happening over and over again was Valiant kept running the king towards the centre to improve its static evaluation score, and was unable to see that it lost tactically.


Fixing that should definitely boost your engine's rating by a few hundred ELO points

If all the piece square tables are backwards, then that ironically means that the archer variant is more tactically correct than standard Valiant, even though Valiant "Archer" has a much lower winrate overall

I have also done some further observations, and to attempt to counter this suicidal royal charge, for the next version I have added tapered piece square tables (albeit, there is no interpolation between the tapering, so in the future I might try to add various types of interpolation).
This has allowed it to play black more successfully, at the cost of losing to 3.2 Standard when playing as white.

In general, it has now been able to beat mildly stronger engines, such as LittleleLouch, Ax 0.8 and BBC 1.4 (with Stockfish NNUE)
Although it does still lose to some strategies, so I would predict it's Elo to be within a blunt range of 1200-2000

[Whiskers]

I got around to fully testing it and I am pretty sure something is wrong with your piece square tables. It keeps running its king up (and getting immediately crushed).
I looked at the eval code to 3.1 (I don't know if these mistakes are also in 3.2, but by its play it seems to be) and found:

1. [rather minor] your queen pst needs to be flipped like all of your other tables
2. [very major!] you are using the king pst for the endgame at all times, including the middlegame! Your search is not very deep, so what kept happening over and over again was Valiant kept running the king towards the centre to improve its static evaluation score, and was unable to see that it lost tactically.


Fixing that should definitely boost your engine's rating by a few hundred ELO points

If all the piece square tables are backwards, then that ironically means that the archer variant is more tactically correct than standard Valiant, even though Valiant "Archer" has a much lower winrate overall

I have also done some further observations, and to attempt to counter this suicidal royal charge, for the next version I have added tapered piece square tables (albeit, there is no interpolation between the tapering, so in the future I might try to add various types of interpolation).
This has allowed it to play black more successfully, at the cost of losing to 3.2 Standard when playing as white.

In general, it has now been able to beat mildly stronger engines, such as LittleleLouch, Ax 0.8 and BBC 1.4 (with Stockfish NNUE)
Although it does still lose to some strategies, so I would predict it's Elo to be within a blunt range of 1200-2000

Have you made sure that all your piece square table lookups are operating properly? For example, if you have a white king on e1, and a black king on e8, does your engine add the correct value for both squares? (can be easily verified with a print statement or two)

Also, an easy way to stop the king from rushing into the middle would be to use this PST from the Simplified evaluation function during the middlegame and only switch to the other one once the amount of material on the board drops below a certain point.

-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-20,-30,-30,-40,-40,-30,-30,-20,
-10,-20,-20,-20,-20,-20,-20,-10,
20, 20, 0, 0, 0, 0, 20, 20,
20, 30, 10, 0, 0, 10, 30, 20

[Enderjed]

I got around to fully testing it and I am pretty sure something is wrong with your piece square tables. It keeps running its king up (and getting immediately crushed).
I looked at the eval code to 3.1 (I don't know if these mistakes are also in 3.2, but by its play it seems to be) and found:

1. [rather minor] your queen pst needs to be flipped like all of your other tables
2. [very major!] you are using the king pst for the endgame at all times, including the middlegame! Your search is not very deep, so what kept happening over and over again was Valiant kept running the king towards the centre to improve its static evaluation score, and was unable to see that it lost tactically.


Fixing that should definitely boost your engine's rating by a few hundred ELO points

If all the piece square tables are backwards, then that ironically means that the archer variant is more tactically correct than standard Valiant, even though Valiant "Archer" has a much lower winrate overall

I have also done some further observations, and to attempt to counter this suicidal royal charge, for the next version I have added tapered piece square tables (albeit, there is no interpolation between the tapering, so in the future I might try to add various types of interpolation).
This has allowed it to play black more successfully, at the cost of losing to 3.2 Standard when playing as white.

In general, it has now been able to beat mildly stronger engines, such as LittleleLouch, Ax 0.8 and BBC 1.4 (with Stockfish NNUE)
Although it does still lose to some strategies, so I would predict it's Elo to be within a blunt range of 1200-2000

Have you made sure that all your piece square table lookups are operating properly? For example, if you have a white king on e1, and a black king on e8, does your engine add the correct value for both squares? (can be easily verified with a print statement or two)

Also, an easy way to stop the king from rushing into the middle would be to use this PST from the Simplified evaluation function during the middlegame and only switch to the other one once the amount of material on the board drops below a certain point.

-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-30,-40,-40,-50,-50,-40,-40,-30,
-20,-30,-30,-40,-40,-30,-30,-20,
-10,-20,-20,-20,-20,-20,-20,-10,
20, 20, 0, 0, 0, 0, 20, 20,
20, 30, 10, 0, 0, 10, 30, 20

Valiant Mk4 does have the same king PST for it's earlygame evaluation. Although it does change to midgame and endgame when turns progress, not when material falls.

[Enderjed]

MK V has finally been released. (along with source)
In terms of changes, the piece evaluation has been renovated, giving the engine a more discernible understanding of the start, mid, and endgame. This is alongside some cosmetic and auditory improvements.
It's strong enough to beat me now in a reasonable timeframe.
Within testing, it's win rate has become far more reliable. So I'd assume it's rating to be closer to 1600-2000.

[Whiskers]

What's Valiant's strongest time control? It kind of struggled when I put it in a 30 seconds + 0.5 sec increment match

Code: Select all

Result:
--------------------------------------------------------------------------
  #  name          games    wins   draws  losses   score    elo    +    -
  1. Celestial       100      98       1       1    98.5    304  105   66
  2. VALIANT MK V    100       1       1      98     1.5   -304   66  105

Cross table:
--------------------------------------------------------------------------
  #  name             score   games                                                                                                    1                                                                                                    2
  1. Celestial         98.5     100                                                                                                    x 111111111111111111111111111111111111111111111111111111=111111111101111111111111111111111111111111111
  2. VALIANT MK V       1.5     100 000000000000000000000000000000000000000000000000000000=000000000010000000000000000000000000000000000                                                                                                    x

Tech:
--------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name            nodes/m         NPS  depth/m   time/m    moves     time   #fails
  1. Celestial          471K      501814      5.9      0.9     30.7     28.8        1
  2. VALIANT MK V         2K        1322      3.2      1.2     30.2     36.4        1
     all ---            232K      222385      4.5      1.1     30.5     32.6        2