Comments/Ratings for a Single Item

I played a game of George Dekle's Hexshogi on this website and easily won. For some reason, it never captured any of my pieces, which quickly gave me a huge lead.

Since a Shogi variant might introduce complexities into the game that could make the AI play more poorly, I tried a game of Chess against Ludii. I easily won that game too. It threw away material and rarely captured my pieces.

As a further test, I conducted a game of Chess between Ludii and Ed Friedlander's Chess applet. The Chess applet did better for most of the game, scooping up lots of material, but it had trouble pressing its advantage in the endgame. It kept on checking the King with its Queen without bringing in more material. The Java Applet declared itself the winner, but it reset the board before reaching checkmate, and I suspect it was a drawn game by the 50 moves rule. Since neither kept track of the moves, I'm not sure how many moves went by in pointless checks.

Since the Java Applet had failed to defeat Ludii at Chess, I conducted another game between the Interactive Diagram for Chess and Ludii. This time, I let Ludii play White to give it the first move advantage. Like in previous games, it threw away its material. However, the Interactive Diagram had trouble pressing its advantage in the endgame, continually checking the King with its Queen instead of bringing in other pieces to support the Queen. Since the Interactive Diagram kept track of the moves, I eventually spotted a three-times repetition, which made the game a draw.

1. e3 d5
2. Bb5 c6
3. Qh5 cxb5
4. d3 Nf6
5. Nc3 Nxh5
6. b4 Nc6
7. e4 Nxb4
8. h4 Nxc2
9. Kd2 Nxa1
10. Rh3 Bxh3
11. a3 dxe4
12. Nxe4 Qa5
13. Ke2 Bxg2
14. Nd6 exd6
15. Bg5 Qxa3
16. f3 Nb3
17. Nh3 Bxh3
18. f4 Nc1
19. Kf2 Nxd3
20. Ke2 Ndxf4
21. Kd1 Kd7
22. Kd2 Qd3
23. Kc1 Kc6
24. Kb2 Qd4
25. Kb3 Be6
26. Kc2 Bf5
27. Kb3 Qc4
28. Kb2 Qb4
29. Ka2 Qc4
30. Kb2 Qb4
31. Ka2 Qc4
32. Kb2 Qb4
33. Ka2 Qc4
34. Kb2 Qb4
35. Ka2 Qc4
36. Kb2 Qb4
37. Ka2 Qc4
38. Kb2 Qb4
39. Ka2 Qc4

It makes sense that Ludii would not play very well. The AI is designed to play a much wider array of games, and it has to sacrifice some performance to do that.

To me — it’s very good site

For another test, I raised how many plies the Interactive Diagram thinks ahead to 4. But while trying to get Ludii to play White, it started playing itself. Since I wanted to see if it could press an advantage in the endgame, I let it continue. Eventually, Black checkmated White. I then ran the planned game with the Interactive Diagram at 4 plies, letting Ludii move first. The Interactive Diagram won this game with checkmate without repeating moves. Here are the moves:

1. e4 e5
2. Qh5 d6
3. Bb5 c6
4. b4 cxb5
5. d4 Nf6
6. Qxf7 Kxf7
7. Ne2 Nxe4
8. d5 Na6
9. Bh6 Nxb4
10. Be3 Nxc2
11. Kf1 Nxa1
12. Bg5 Qxg5
13. Nf4 Qxf4
14. f3 Qc1
15. Ke2 Qxh1
16. fxe4 Qxb1
17. Ke3 Qxa2
18. h3 Qxg2
19. h4 Qh3
20. Kd2 Qxh4
21. Kc3 Qxe4
22. Kb2 Nc2
23. Kc3 Bf5
24. Kb2 Qxd5
25. Kc1 Qd4
26. Kb1 Qa1

Ludii is interesting, but the language it uses for defining games confuses me so much I can't imagine trying to describe anything as complex as a chess variant with it.

I saved the PDF on the Ludii language to my Kindle DX to review later.

However, the Interactive Diagram had trouble pressing its advantage in the endgame, continually checking the King with its Queen instead of bringing in other pieces to support the Queen.

The Interactive Diagram is rather poor in end-games. I am aware of that, but did not consider it a high priority to improve. The purpose of the Diagram is to give the user a good impression of what a given chess variant is like, not to score as high as possible in a tournament. An end-game is often very different in character than the game it results from. For getting an idea of how to win with a specific piece the Checkmating Applets are a convenient tool.

The most useful definition of end-game that I know is when it is no longer fatal to involve your royal piece in the battle, and let it roam the board. Because royal pieces tend to have very limited power of motion, this requires planning many moves ahead, and the key positions where progress can be made in terms of gaining material can usually be reached through many different paths. To find those requires lare search depths, which can only be reached with the aid of a Transposition Table. With the aid of that you can avoid having to evaluate the same position over and over again depending on the path you used to reach it. The AI of the Interactive Diagram does not have a transposition table (yet). So the search depth can not be made large enough to allow it to see the ways to make progress, which makes it clueless. Middle-game tactics, on the other hand, is quite shallow in comparison, and transpositions are not very abundant, so not having a TT is not much of a drawback.

Chasing the royal piece with checks is a problem that can be prevented even without a TT, though. The trick is that it is in general useful to check. (It limits the number of moves available to the opponent, and who knows what opportunity will present itself after that?) But repetitive checking with the same piece in general leads to nothing, except when the piece has mating power on its own (such as the Lion, KNAD). So one can (lightly) penalize branches that contain multiple checks with the same piece in a row. While checking with two different pieces usually leads to a quick checkmate.

Problem is that currently the Interactive Diagram is not even aware when it is in check, except in Quiescence Search (where it then suppresses scoring the position with the current material balance, and forces search of all possible check evasions), or when it could not find any legal move. But at some point I will probably add some in-check detection, and then I can add this 'anti-chasing' heuristic too.

I downloaded the Ludii player to my computer and ran a game against the Interactive Diagram at its default settings. This time, the Interactive Diagram quickly won with checkmate. I wonder if it makes a difference that I was using Edge instead of Firefox, given that I keep Firefox loaded with tabs, and I always start Edge fresh with no old tabs open.

1. d4 e5
2. dxe5 Qg5
3. Bxg5 Bb4
4. c3 f6
5. exf6 Bxc3
6. Nxc3 d5
7. fxg7 h6
8. gxh8=Q Nc6
9. Qxg8 Kd7
10. Qdxd5

Since the Ludii player allows me to adjust the thinking time, I will see if additional thinking time will allow it to defeat the Interactive Diagram.

In the previous game, the Ludii player had 1 second of thinking time. I next tried it with 5 seconds, but when it was about to throw away a Queen, I started a new game with 10 seconds of thinking time for it. The Interactive Diagram played in Edge with its default settings. In the start of the game, Ludii moved out its Queen as it had done at 5 seconds, but instead of throwing it away, it fled from attacks, letting its Queen get chased around. But it eventually began to lose material, and by the endgame, it had only the King left. The Interactive Diagram eventually defeated it with checkmate. Here is the game:

1. d4 e6
2. c3 Qh4
3. Nf3 Qe4
4. Nbd2 Qf4
5. Nb3 Qe4
6. Nc5 Bxc5
7. dxc5 Kf8
8. Qd4 Qxd4
9. Nxd4 Nf6
10. Bf4 Ke8
11. Bxc7 Ne4
12. Be5 d6
13. Bxg7 e5
14. Bxh8 exd4
15. cxd4 Bg4
16. cxd6 Nd7
17. h3 Be6
18. e3 Rc8
19. Bd3 f6
20. Bxe4 Bf7
21. Bxb7 Rc2
22. b3 Be6
23. O-O Kf7
24. Be4 Rb2
25. Bxh7 Bd5
26. Bf5 Rd2
27. Bxd7 Bxg2
28. Kxg2 Re2
29. Kf3 Rd2
30. Bf5 Rb2
31. Ke4 Rc2
32. Kd3 Rc6
33. d7 a6
34. d8=Q Re6
35. Bxe6 Kxe6
36. Qxf6 Kd7
37. Qxa6 Kc7
38. Be5 Kd7
39. Qb5 Ke7
40. Ke4 Kf7
41. Qd5 Kg6
42. Qg8 Kh5
43. Qg4 Kh6
44. Kf5 Kh7
45. Qg6

For the next game, I gave the Ludii player 30 seconds of thinking time. It still played very poorly, making lots of obviously bad moves. The Interactive Diagram made sensible moves for most of the game, but during the endgame, it failed to mobilize its overwhelmingly superior forces for checkmate. It could have easily gotten checkmate by moving out a Rook to help the Queen, but instead it moved out its King. When it was getting close to further hemming in the King and eventually getting checkmate, it moved its Queen to a rank further from the King, giving it more space to move. Eventually, the game ended in a draw through three-times repetition, which I estimated by observing the same set of four moves repeat three times. Here are the moves:

1. d4 e6
2. c3 Qh4
3. Nf3 Qe4
4. Nbd2 Qf4
5. Nb3 Qe4
6. Nc5 Bxc5
7. dxc5 d6
8. cxd6 Bd7
9. dxc7 Ne7
10. cxb8=Q Rxb8
11. Be3 O-O
12. Qxd7 Qc4
13. Qxe7 Rbd8
14. Qxb7 Rd6
15. Qxa7 e5
16. Nxe5 Qh4
17. g3 Qe7
18. Qxe7 Rd5
19. Bg2 Rxe5
20. Qxe5 Rd8
21. O-O h6
22. h4 h5
23. Qxh5 f6
24. Bd5 Rxd5
25. Qxd5 Kh7
26. Qh5 Kg8
27. Kg2 Kf8
28. Kf3 g6
29. Qxg6 Ke7
30. Bc5 Kd8
31. Ke4 Kc7
32. Qxf6 Kb8
33. Kd5 Kc8
34. Qf5 Kc7
35. Qf4 Kc8
36. Qf5 Kc7
37. Qf4 Kc8
38. Qf5 Kc7
39. Qf4 Kc8
40. Qf5

Well, the default setting ('2.5 ply') of the ID is quite weak; it already has difficulty in recognizing mate-in-1 threats there. In an end-game with just a few pieces it would play very fast. An AI that is controlled by time rather than depth would automatically start to search deeper in the end-game. Humans would think farther ahead in a simple end-game than in a complex middle-game, so with a fixed search depth at some point in the game the AI will start to look stupid, even when it initially would be percieved as a tough opponent that knows what it is doing. I could of course control the playing strength through specifying seconds, or a number of nodes, instead of ply. The advantage of ply is that it provides an absolute measure for the tactical capabilities (e.g. how far ahead it will recognize a checkmate), independent of the complexity of the variant.

To seriously compare the strength to that of an AI controlled by thinking time it would be necessary to keep the time used by the ID (which it reports after every move above the board) approximately at the same level of that of its opponent during the game. Which would require cranking up the depth as the game simplifies.

It seems that the Ludii AI is simply buggy. A good search should not make blunders of the kind it does, even if it is not especially optimized for Chess variants. Jocly is also uses a very general (MCTS) search, because of its general scope. But I have never seen it make blunders like this. Ai Ai is quite prone to blundering, though.

Which ai are you using in ludii?

Which ai are you using in ludii?

Ludii.

There is also an alpha-beta ai which is stronger for some games

Since I figured out how to make the Ludii Player show the last move, I decided to try one more game with it set to its maximum thinking time of 300 seconds, which is 5 minutes. As before, it moved out its Queen and let it get chased around a bit. Overall, it played very poorly, and it was checkmated in 32 moves, making this game even shorter than previous games with less thinking time for Ludii. In the endgame, it seemed to just give up, letting checkmate come even more quickly than it had to. Based on this, I would conclude that Ludii is not nearly as good at playing Chess variants as the Interactive Diagram is. And since that is itself a low bar, my conclusion is that Ludii is not very good for Chess variants. Here's the game:

1. d4 e6
2. c3 Qh4
3. Nf3 Qe4
4. Nbd2 Qd5
5. e4 Qh5
6. Bc4 b5
7. Bd3 Qg4
8. O-O Nc6
9. Bxb5 Bb4
10. cxb4 Nge7
11. h3 Qf4
12. g3 Qd6
13. Nc4 Qxb4
14. Qb3 O-O
15. Qxb4 Nxb4
16. Bf4 Bb7
17. Nfd2 Kh8
18. Bxd7 Rad8
19. Ne5 f6
20. Bxe6 fxe5
21. Bxe5 Nd3
22. Bxc7 Rxd4
23. Ba5 Bc6
24. Bc3 Ra4
25. Bb3 Nd5
26. exd5 Rf5
27. dxc6 Rg5
28. Bxa4 Rh5
29. h4 Ne5
30. Bb4 Rh6
31. c7 Rh5
32. c8=Q

There is also an alpha-beta ai which is stronger for some games

I just ran a game between Ludii as White and Alpha-Beta as Black at 1sec each, and Ludii quickly won. Here's what it reported the moves to be:

1. (Move E2-E4)
2. (Move E7-E5)
3. (Move F1-B5)
4. (Move D8-H4)
5. (Move D2-D4)
6. (Move F8-B4)
7. (Move B1-D2)
8. (Remove F2)
9. (Remove F2)
10. (Remove D2)
11. (Remove D2)
12. (Move B8-C6)
13. (Move D1-H5)
14. (Remove D4)
15. (Remove E5)
16. (Move G8-E7)
17. (Remove D4)
18. (Move E7-D5)
19. (Remove D5)
20. (Move C7-C5)
21. (Move D4-E5)
22. (Move E8-F8)
23. (Move G1-F3)
24. (Move D7-D6)
25. (Move E5-E8)

Ludii as White beat Random as Black in even fewer moves:

1. (Move E2-E4)
2. (Move G8-H6)
3. (Move F1-C4)
4. (Move H6-G8)
5. (Remove F7)
6. (Remove F7)
7. (Move D1-H5)
8. (Move F7-E6)
9. (Move H5-F5)
10. (Move E6-D6)
11. (Move F5-D5)

At 1sec per move, Ludii (as White) beat Flat MC (as Black) in 99 moves:

1. (Move E2-E4)
2. (Move C7-C6)
3. (Move D1-H5)
4. (Move D8-C7)
5. (Move F2-F4)
6. (Move D7-D5)
7. (Move F1-A6)
8. (Move C8-E6)
9. (Move A2-A4)
10. (Move E6-C8)
11. (Move D2-D4)
12. (Remove E4)
13. (Move B2-B4)
14. (Move E7-E6)
15. (Move B1-C3)
16. (Move G8-H6)
17. (Move C1-B2)
18. (Move F8-E7)
19. (Remove F7)
20. (Remove F7)
21. (Move A6-C4)
22. (Remove B4)
23. (Move E1-E2)
24. (Move F7-H6)
25. (Move G1-H3)
26. (Move E8-D8)
27. (Move A4-A5)
28. (Move C7-B6)
29. (Remove B6)
30. (Move D8-E8)
31. (Remove E4)
32. (Move A7-A6)
33. (Move H3-G5)
34. (Move G7-G6)
35. (Move E4-F6)
36. (Move E8-F8)
37. (Move A1-A3)
38. (Move H8-G8)
39. (Move H1-D1)
40. (Move A8-A7)
41. (Move H2-H4)
42. (Move B4-E7)
43. (Remove H7)
44. (Move F8-G7)
45. (Move A3-A5)
46. (Move H6-G4)
47. (Remove G4)
48. (Move E7-D8)
49. (Move D1-F1)
50. (Move G7-F7)
51. (Move G4-H6)
52. (Move F7-G7)
53. (Remove G8)
54. (Move G7-F7)
55. (Move G8-H6)
56. (Move F7-E7)
57. (Move H6-G8)
58. (Move E7-F7)
59. (Move H7-F8)
60. (Move D8-F6)
61. (Remove F6)
62. (Move A7-A8)
63. (Move F8-H7)
64. (Move C6-C5)
65. (Move E2-F2)
66. (Move G6-G5)
67. (Move F1-G1)
68. (Move C8-D7)
69. (Move G2-G4)
70. (Move D7-E8)
71. (Move A5-A3)
72. (Move E8-D7)
73. (Move F6-E4)
74. (Move D7-C6)
75. (Remove G5)
76. (Move F7-E8)
77. (Move D4-D5)
78. (Remove D5)
79. (Move G1-H1)
80. (Move B8-C6)
81. (Move H4-H5)
82. (Move E8-D8)
83. (Move G5-F7)
84. (Move D8-D7)
85. (Remove C5)
86. (Move D7-C8)
87. (Move F7-D6)
88. (Move C8-D8)
89. (Move B2-F6)
90. (Move C6-E7)
91. (Remove D5)
92. (Move A8-C8)
93. (Remove C8)
94. (Move A6-A5)
95. (Move C8-D6)
96. (Remove D5)
97. (Move H1-E1)
98. (Move D5-D4)
99. (Remove E7)

Since the Ludii Player has many engines (or agents) to choose from, I had it do an analysis to find the best player for Chess. The analysis was quick, and it said "Best predicted agent is AlphaBeta". But I already found that it played poorly against Ludii. In case more thinking time would eventually give Alpha-Beta an edge, I increased the thinking time to 5 seconds for another game between Ludii as White and Alpha-Beta as Black. Ludii won this game with checkmate. During the game, it showed an analysis of each engine, and Ludii, identified here as UBFM consistently had a higher search depth than Alpha-Beta.

UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.13999999,
1196 different states were evaluated
50 iterations, with 134 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.20000005,
909 different states were evaluated
30 iterations, with 68 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.54125,
1014 different states were evaluated
33 iterations, with 97 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.6924999,
1002 different states were evaluated
28 iterations, with 75 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 6:
best value observed at root 1.2237501,
837 different states were evaluated
32 iterations, with 110 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.7437501,
878 different states were evaluated
37 iterations, with 114 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.66375005,
881 different states were evaluated
36 iterations, with 72 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.7237501,
883 different states were evaluated
35 iterations, with 101 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 2:
best value observed at root 0.6237501,
894 different states were evaluated
33 iterations, with 65 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.9375,
908 different states were evaluated
24 iterations, with 65 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.9112501,
918 different states were evaluated
25 iterations, with 63 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.9512501,
1025 different states were evaluated
39 iterations, with 80 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.77125,
900 different states were evaluated
35 iterations, with 71 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 2:
best value observed at root 0.9112501,
909 different states were evaluated
34 iterations, with 67 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.73125005,
912 different states were evaluated
34 iterations, with 69 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 2:
best value observed at root 0.6712501,
794 different states were evaluated
22 iterations, with 43 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.68000007,
803 different states were evaluated
20 iterations, with 55 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.6575,
875 different states were evaluated
23 iterations, with 50 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 0.87125003,
920 different states were evaluated
22 iterations, with 60 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.90250015,
1043 different states were evaluated
26 iterations, with 70 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 1.15625,
1223 different states were evaluated
32 iterations, with 98 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.8962499,
1111 different states were evaluated
44 iterations, with 88 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 0.9849999,
1272 different states were evaluated
51 iterations, with 108 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 1.2962499,
1107 different states were evaluated
26 iterations, with 66 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 1.0362499,
1006 different states were evaluated
32 iterations, with 65 calls of minimax
Alpha-Beta (player 2) completed search of depth 1.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 1.7525,
1297 different states were evaluated
35 iterations, with 100 calls of minimax
Alpha-Beta (player 2) completed search of depth 2.
UBFM (player 1) completed an analysis that reached at some point a depth of 5:
best value observed at root 1.6012498,
1267 different states were evaluated
49 iterations, with 187 calls of minimax
Alpha-Beta (player 2) completed search of depth 3.
UBFM (player 1) completed an analysis that reached at some point a depth of 3:
best value observed at root 1.68125,
1304 different states were evaluated
41 iterations, with 99 calls of minimax
Alpha-Beta (player 2) completed search of depth 3.
UBFM (player 1) completed an analysis that reached at some point a depth of 9:
best value observed at root 1.6675,
1278 different states were evaluated
101 iterations, with 398 calls of minimax
Alpha-Beta (player 2) completed search of depth 3.
UBFM (player 1) completed an analysis that reached at some point a depth of 7:
best value observed at root 1.5674999,
1288 different states were evaluated
107 iterations, with 429 calls of minimax
Alpha-Beta (player 2) completed search of depth 3.
UBFM (player 1) completed an analysis that reached at some point a depth of 5:
best value observed at root 1000000.0,
954 different states were evaluated
323 iterations, with 1225 calls of minimax
Alpha-Beta (player 2) completed search of depth 0.
UBFM (player 1) completed an analysis that reached at some point a depth of 4:
best value observed at root 1000000.0,
1237 different states were evaluated
753 iterations, with 1541 calls of minimax

Here is the game:

1. (Move E2-E4)
2. (Move E7-E5)
3. (Move D1-H5)
4. (Move D8-H4)
5. (Remove H4)
6. (Move G7-G5)
7. (Remove G5)
8. (Move D7-D5)
9. (Move F1-B5)
10. (Move C7-C6)
11. (Remove E5)
12. (Move E8-D8)
13. (Move E5-G5)
14. (Move D8-C7)
15. (Move G5-E5)
16. (Move C7-B6)
17. (Move B5-E2)
18. (Move C8-G4)
19. (Remove G4)
20. (Move B8-D7)
21. (Remove D7)
22. (Move B6-A6)
23. (Move G1-E2)
24. (Move F8-B4)
25. (Move B2-B3)
26. (Move F7-F5)
27. (Move A2-A3)
28. (Move B4-C5)
29. (Move A1-A2)
30. (Move G8-F6)
31. (Move C1-B2)
32. (Remove F2)
33. (Move E1-D1)
34. (Remove D7)
35. (Remove F5)
36. (Move H8-G8)
37. (Remove F2)
38. (Move D5-D4)
39. (Remove D4)
40. (Move D7-C5)
41. (Remove C5)
42. (Move A8-D8)
43. (Move B3-B4)
44. (Move B7-B5)
45. (Move D1-C1)
46. (Move D8-D7)
47. (Move F2-F6)
48. (Move A6-B7)
49. (Move H1-D1)
50. (Move D7-D4)
51. (Remove D4)
52. (Remove G2)
53. (Move F6-F7)
54. (Move B7-C8)
55. (Move D1-F1)
56. (Remove E2)
57. (Move F7-G8)
58. (Move C8-D7)
59. (Move F1-F7)
60. (Move D7-D6)
61. (Move G8-D8)
62. (Move D6-E6)
63. (Move F7-F6)

I have been running one game after another on Ludii Player between Ludii as White and another agent as Black at one second thinking time each. Going from top down, each has lost to Ludii until I got to EPT, which beat Ludii. Here is that game:

Turn 1. E2-E4
Turn 2. E7-E6
Turn 3. B2-B4
Turn 4. B4-
Turn 5. C1-B2
Turn 6. D7-D5
Turn 7. F1-B5
Turn 8. E8-F8
Turn 9. D1-G4
Turn 10. G7-G6
Turn 11. B2-G7
Turn 12. G7-
Turn 13. G4-G5
Turn 14. G5-
Turn 15. G1-E2
Turn 16. C8-D7
Turn 17. D7-
Turn 18. D7-
Turn 19. E1=0
Turn 20. B4-F8
Turn 21. G2-G4
Turn 22. C7-C6
Turn 23. G1-G2
Turn 24. G5-H4
Turn 25. B1-C3
Turn 26. F2-
Turn 27. F2-
Turn 28. F7-F5
Turn 29. D5-
Turn 30. F8-B4
Turn 31. B4-
Turn 32. G8-F6
Turn 33. F2-F3
Turn 34. A8-B8
Turn 35. A2-A4
Turn 36. H8-G8
Turn 37. G4-G5
Turn 38. E4-
Turn 39. A4-A5
Turn 40. E4-G3
Turn 41. G3-
Turn 42. D7-E5
Turn 43. B4-D5
Turn 44. F3-
Turn 45. F3-
Turn 46. D5-
Turn 47. A1-B1
Turn 48. B8-C8
Turn 49. B7-
Turn 50. C8-C7
Turn 51. C7-
Turn 52. G7-H8
Turn 53. E2-D4
Turn 54. G8-G7
Turn 55. C7-C8
Turn 56. G7-G8
Turn 57. C8-B8
Turn 58. B8-
Turn 59. F3-F2
Turn 60. B8-B2
Turn 61. E6-
Turn 62. C2-
Turn 63. E6-D4
Turn 64. D2-
Turn 65. D4-E2
Turn 66. A7-A6
Turn 67. H2-H4
Turn 68. E2-
Turn 69. E2-
Turn 70. H8-G8
Turn 71. E2-D2
Turn 72. G8-F7
Turn 73. D2-C2
Turn 74. F7-E6
Turn 75. C2-C3
Turn 76. E6-D6
Turn 77. C3-B2
Turn 78. D5-D4
Turn 79. B2-B3
Turn 80. D6-D5
Turn 81. B3-B2
Turn 82. D5-C4
Turn 83. B2-C2
Turn 84. D4-D3
Turn 85. C2-B2
Turn 86. D3-D2
Turn 87. B2-C2
Turn 88. D2-D1, D1 => Bishop2
Turn 89. D1-
Turn 90. F5-F4
Turn 91. D1-C2
Turn 92. F4-F3
Turn 93. C2-B2
Turn 94. F3-F2
Turn 95. B2-C2
Turn 96. F2-F1, F1 => Queen2
Turn 97. H4-H5
Turn 98. F1-F4
Turn 99. C2-B1
Turn 100. G5-
Turn 101. H5-H6
Turn 102. C4-B3
Turn 103. B1-A1
Turn 104. G5-C1

I'm done for the night.

The reported search depths for alpha-beta are absurdly low. Most engines would reach such depths faster than a millisecond. So there doesn't seem to be a reason to stop there. In one place it even reports a depth of 0. Either it lies about its depth, or ther is something very wrong with this alpha-beta implementation.