-
Notifications
You must be signed in to change notification settings - Fork 0
/
log.txt
1045 lines (899 loc) · 42.4 KB
/
log.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
_____________________________________
Captain's Log 6/25/21
Pushing to github...
The current neural network doesn't seperate the features at all
(features are just merged into a layer before they are reduced at all).
I just fixed major issues in both board generation and in the application
of MCTS (turns out we had to pick the *minimum* board, since we're playing
hypothetical games for the *opponent*). I also fixed an issue with testing:
I was flipping the turn order of strategies 1 and 2 when they are pitted
against each other, but I needed to *test* when order matters.
The parameters of the net are:
self.parameters = {
"layer1 activ func" : 'gelu',
"layer2 activ func" : 'gelu',
"output activ func" : 'linear',
"loss func" : 'mean_squared_error',
"num epochs" : 1000000000, # Note that we *really* stop when the net converges
"shuffle" : True,
"batch size" : 32
}
Training results:
2x2:
loss 0.0022, accuracy 100.0%
Against random
P1 100.0% P2 0.0%
Against itself
P1 100.0% P2 0.0%
3x3:
loss 0.0098, accuracy 75.43%
Against random
P1 95.0% P2 80.0%
Against itself
P1 0.0% P2 100.0%
4x4:
loss 0.0465, accuracy 29.71%
Against random
P1 95.0% P2 92.5%
Against itself
P1 0.0% P2 100.0%
5x5:
loss 0.0291, accuracy 17.80%
Against random
P1 55.0% P2 60.0%
Against itself
P1 0.0% P2 100.0%
UPDATE: 6/29/21
Separating the features isn't really necessary in this network; a "plain-ol"
neural network might be enough.
This I noticed because of the following results about the above parameters
with this net:
input_A = Input(shape=(1,))
input_B = Input(shape=(1,))
input_C = Input(shape=(2*n + 2,))
input_D = Input(shape=(n**2,))
tempo_layer = Dense(1, activation=self.layer1actfn)(input_A)
dead_squares_layer = Dense(1, activation=self.layer1actfn)(input_B)
parity_layer = Dense(int(2*n+2), activation=self.layer1actfn)(input_C)
positional_layer = Dense(int(n**2), activation=self.layer1actfn)(input_D)
# We merge the features, and then include layers for high-level feature extraction and output.
merged_layer = Concatenate()([tempo_layer, dead_squares_layer, parity_layer, positional_layer])
abstract_layer = Dense(int(n**2), activation=self.layer2actfn)(merged_layer)
output_layer = Dense(1, activation=self.layer_out_actfn)(abstract_layer)
self.model = tf.keras.Model(inputs=[input_A, input_B, input_C, input_D], outputs=output_layer)
Training results:
2x2: (10 boards)
loss 0.0023, accuracy 100.0%
Against random
P1 100.0% P2 0.0%
Against itself
P1 100.0% P2 0.0%
3x3: (436 boards)
loss 0.0111, accuracy 75.0%
Against random
P1 91.0% P2 87.0%
Against itself
P1 0.0% P2 100.0%
4x4: (13962 boards)
loss 0.0301, accuracy 32.10%
Against random
P1 94.0% P2 91.0%
Against itself
P1 0.0% P2 100.0%
5x5: (57836 boards)
loss 0.0289, accuracy 18.20%
Against random
P1 74.0% P2 46.0%
Against itself
P1 100.0% P2 0.0%
*Note: "against itself" data is probably mostly trash,
given that only one unique game is ever played.
UPDATE: 9/2/21
I'm running out of time. So at this point, I'm going to systematically test the following
choices against each other on a reasonable size game (4x4 boards):
1. feature selection
2. architecture choice
3. generating boards
(These nets were trained with patience=20, min_delta=0.0001 early stopping,
and were tested against random on 20 4x4 boards.)
KEY:
init->nbhd->flipped : A procedure for generating boards.
First produces some initial boards, then gathers boards in its neighborhood,
then flips whose turn it is on these boards. Just for board generation,
the role of the teacher is de-emphasized.
modal : A network architecture. Multi-layered; in the first layer, features of different
'modalities' (e.g. 'global' vs 'local' features) are separated, and then combined
flat : A network architecture. All the features are fed in, but there is a single
"flattened" layer that abstracts away features before the output.
____________________________________________________________________________________________________________
TEST NO _
____________________________________________________________________________________________________________
FEATURES: _
ARCHITECTURE: _
BOARD GEN: _
_______________
4x4: (_ boards)
loss (mse) _
Against random
P1 _% P2 _%
____________________________________________________________________________________________________________
############################################################################################################
9/2/21 Comparing and contrasting feature choices
____________________________________________________________________________________________________________
TEST NO 1
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of taken, dead, available squares; number of taken, dead, available squares
ARCHITECTURE: modal
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0103
Against random
P1 75% P2 65%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 2
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of taken, dead, available squares; number of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0374
Against random
P1 75% P2 60%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 3
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken, dead, available squares
ARCHITECTURE: modal
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0088
Against random
P1 60% P2 75%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 4
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0323
Against random
P1 75% P2 70%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 5 : trying to replicate 6/29/21 (unsuccessfully)
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0717
Against random
P1 85% P2 50%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 6
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (available); map of squares (taken)
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0777
Against random
P1 55% P2 65%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 7 : no 4, but with n**2 layer dimension instead of 2(n**2)
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0459
Against random
P1 80% P2 75%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 8 : no 7, but min_delta=0.005 instead of 0.0001, since I suspect the network is overfitting
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0574
Against random
P1 50% P2 55%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 9 : no 7, but min_delta=0.00005 instead of 0.0001, patience=30 since I suspect the network is underfitting
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0456
Against random
P1 55% P2 70%
____________________________________________________________________________________________________________
Thoughts:
- Test no. 7 is the best, but it's nothing to sneeze at
- Why reinvent the wheel? Instead, I think I should:
- Take an implementation of AlphaZero
- Play around with the features, net architecture, and board generation method within it
- This also allows me to compare my results against Saul's paper as a baseline.
- See
https://github.com/suragnair/alpha-zero-general
for general game-playing with AlphaZero.
############################################################################################################
9/6/21 Convolutional Net Tests
____________________________________________________________________________________________________________
TEST NO 10: Transferred AlphaZero's convolutional net
____________________________________________________________________________________________________________
FEATURES: map of taken
ARCHITECTURE: convolutional
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0965
Against random
P1 80% P2 60%
4x4: (13962 boards)
loss (mse) 0.0126
Against random
P1 50% P2 60%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 11: Adding features to the convolutional net (hybrid of tests 7 and 10)
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map) feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards boards)
loss (mse) 0.0141
Against random
P1 85% P2 85%
4x4: (13962 boards)
loss (mse) _
Against random
P1 _% P2 _%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 12: Convolutional net + just the tempo flag
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of squares (taken)
ARCHITECTURE: convolutional (map) feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse)
Against random
P1 60% P2 55%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 13: Replication of the 6/19/21 net!!!
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) _
Against random
P1 85% P2 95%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 14: 6/19/21, but without num of dead squares
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0298
Against random
P1 95% P2 85%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 15: 6/19/21, but with only tempo flag
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0819
Against random
P1 65% P2 55%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 16: 6/19/21, but with only parity information
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0297
Against random
P1 85% P2 90%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 17: Convolutional net + features from test 14
____________________________________________________________________________________________________________
FEATURES: tempo flag; parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: convolutional (map) feeding into flat (using a merged layer)
BOARD GEN: init->nbhd->flipped
_______________
4x4: (_ boards)
loss (mse) 0.0046
Against random
P1 65% P2 60%
____________________________________________________________________________________________________________
Thoughts:
(Maybe) Why AlphaZero performs poorly on Notakto;
In all games AlphaZero performs well on, e.g. go, chess, checkers, tic-tac-toe, and connect 4,
it's easy to determine who the winner is by the final board state. e.g. in chess see whose king
is cornered, in checkers see whose pieces remain, in tic-tac-toe see whether X or O has n in a row.
But in Notakto, the winner is ambiguous from the final board state alone. Since all players play
'X', the neural network is not given enough information to form a pattern between boards and winners.
So the net needs to be fed this 'tempo bit' to tell it whose turn it is in each state.
Another potential hangup is that Notakto uses misere play -- in all other AlphaZero games, the final
move by a player gives them the win. In Notakto, the final move by a player gives them a loss.
This might be "negative information" that is difficult for a neural network to learn from the
board states alone. (I'm unsure how exactly this affects the learning process, though)
Update 9/8: Yet another issue with AlphaZero might be that it uses a convolutional neural network.
A convolutional net forms relationships between a square and surrounding squares in its region.
But in Notakto, surrounding squares are not all necessarily relevant, leading the convolutional net
to go down bad search trees.
For example, in the board
_______
|_|_|_|_|
|_|_|_|B|
|_|_|A|_|
|_|_|_|_|
Square B is immediately diagonal of square A. But because B is not on the same row, column, or diagonal
as A, an X on B has only indirect bearing on whether an X should be placed at A. This is dissimilar
to chess, where a pawn or bishop may attack diagonally (and so it would be important to track this
information).
All of this demonstrates just how much domain knowledge AlphaZero is really assuming about the game
it plays. Unfortunately, any choice in features whatsoever reflects knowledge about one's domain
-- what is important, what is not -- and AlphaZero's convolutional net is no exception.
############################################################################################################
9/8/21 4x4, Trying to improve on best
CONTROL GROUP:
____________________________________________________________________________________________________________
TEST NO 10: Transferred AlphaZero's convolutional net
____________________________________________________________________________________________________________
FEATURES: map of taken
ARCHITECTURE: convolutional
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0965
Against random
P1 80% P2 60%
4x4: (13962 boards)
loss (mse) 0.0126
Against random
P1 50% P2 60%
____________________________________________________________________________________________________________
BEST BOARD FROM PREVIOUS:
____________________________________________________________________________________________________________
TEST NO 13: 6/19/21 (aka Test 13)
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) _
Against random
P1 85% P2 95%
____________________________________________________________________________________________________________
(The following nets were trained with patience=20, min_delta=0.0001 early stopping,
and were tested against random on 100 4x4 boards.)
____________________________________________________________________________________________________________
TEST NO _
____________________________________________________________________________________________________________
FEATURES: _
ARCHITECTURE: _
BOARD GEN: _
_______________
4x4: (13962 boards)
loss (mse) _
Against random
P1 _% P2 _%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 18: CONTROL GROUP: AlphaZero convolutional net
____________________________________________________________________________________________________________
FEATURES: map of taken
ARCHITECTURE: convolutional
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) _
Against random
P1 _% P2 _%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 19: PREVIOUS BEST: 6/19/21 (aka Test 13)
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0237
Against random
P1 93% P2 93%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 20:
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row/col/diag (taken); map of squares (taken) (row/col/diag put in independently)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
4x4: (13962 boards)
loss (mse) 0.0131
Against random
P1 92% P2 92%
____________________________________________________________________________________________________________
---------------------------------------------------------
Official Tests
Systems:
Conv Net - AlphaZero's Convolutional NN (adapted from TicTacToe)
Feature Net - Our winner, system from Test 19
Board sizes: 3x3, 4x4, 5x5, 6x6
Early stopping:
min_delta=0.0001
patience=20
Against:
random_selection
greedy
We cannot test against itself, since this prediction process is
entirely deterministic (and so the games would end up being exactly
the same). See:
https://chess.stackexchange.com/questions/19482/randomness-in-engine-play
AlphaZero uses actual MCTS (our tree search in the preliminary tests isn't
actually random at all), and so AlphaZero will play a different game each time.
---------------------------------------------------------
PRELIMINARY TESTS (basic MCTS, learn from training set)
---------------------------------------------------------
---------------------------- ---------------------------
CONV NET FEATURE NET
---------------------------- ---------------------------
3x3: (436 boards) 3x3: (436 boards)
loss (mse) _.___ loss (mse) 0.0019
Against random Against random
P1 __% P2 __% P1 89% P2 84%
Against greedy Against greedy
P1 __% P2 __% P1 72% P2 46%
Against itself Against itself
P1 __% P2 __% P1 61% P2 39%
4x4: (13962 boards) 4x4: (13962 boards)
loss (mse) _.___ loss (mse) 0.0180
Against random Against random
P1 __% P2 __% P1 94% P2 95%
Against greedy Against greedy
P1 __% P2 __% P1 39% P2 58%
Against itself Against itself
P1 __% P2 __% P1 39% P2 61%
5x5: ( boards) 5x5: ( boards)
loss (mse) _.___ loss (mse) _.___
Against random Against random
P1 __% P2 __% P1 __% P2 __%
Against greedy Against greedy
P1 __% P2 __% P1 __% P2 __%
Against itself Against itself
P1 __% P2 __% P1 __% P2 __%
6x6: ( boards) 6x6: ( boards)
loss (mse) _.___ loss (mse) _.___
Against random Against random
P1 __% P2 __% P1 __% P2 __%
Against greedy Against greedy
P1 __% P2 __% P1 __% P2 __%
Against itself Against itself
P1 __% P2 __% P1 __% P2 __%
---------------------------------------------------------
ALPHAZERO TESTS (learn via repeated self-play)
---------------------------------------------------------
Update: 9/10/2021
I realize now that the goal should not be competent play
against random or greedy. The goal should be to have the
neural network *learn the strategy*, which we can see by its
self-play results. So I'm going to focus now *strictly* on
getting a net to successfully learn the strategy for a 3x3
board.
____________________________________________________________________________________________________________
TEST NO __:
____________________________________________________________________________________________________________
FEATURES:
ARCHITECTURE:
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) _.____
Against itself
P1 __% P2 __%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 21:
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row/col/diag (taken); map of squares (taken) (row/col/diag put in independently)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0019
Against itself
P1 60% P2 40%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 22:
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of taken, dead, available squares; number of taken, dead, available squares
ARCHITECTURE: modal
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0009
Against itself
P1 65% P2 35%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 23:
____________________________________________________________________________________________________________
FEATURES: tempo flag; map of taken, dead, available squares; number of taken, dead, available squares
ARCHITECTURE: flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0147
Against itself
P1 57% P2 43%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 24:
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0177
Against itself
P1 65% P2 35%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 26:
____________________________________________________________________________________________________________
FEATURES: map of taken
ARCHITECTURE: convolutional
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0479
Against itself
P1 77% P2 23%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 27:
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0390
Against itself
P1 68% P2 32%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 28: Same as test 25, but abstraction layer has size n instead of n**2
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (436 boards)
loss (mse) 0.0390
Against itself
P1 85% P2 15%
____________________________________________________________________________________________________________
--------------------------------------------------------------
PRELIMINARY TESTS (basic tree search, learn from training set)
--------------------------------------------------------------
---------------------------- -------------------------------------------
CONV NET CONV + FEATURES + ABSTRACT LAYER (size n)
---------------------------- -------------------------------------------
3x3: (436 boards) 3x3: (450 boards)
loss (mse) 0.1035 loss (mse) 0.0803
Against random Against random
P1 91% P2 86% P1 91% P2 85%
Against greedy Against greedy
P1 74% P2 25% P1 69% P2 43%
********************* *********************
Against itself Against itself
P1 100% P2 0% P1 85% P2 15%
********************* *********************
4x4: (13962 boards) 4x4: (13962 boards)
loss (mse) _.___ loss (mse) _.____
Against random Against random
P1 __% P2 __% P1 __% P2 __%
Against greedy Against greedy
P1 __% P2 __% P1 __% P2 __%
********************* *********************
Against itself Against itself
P1 __% P2 __% P1 __% P2 __%
********************* *********************
5x5: ( boards) 5x5: ( boards)
loss (mse) _.___ loss (mse) _.___
Against random Against random
P1 __% P2 __% P1 __% P2 __%
Against greedy Against greedy
P1 __% P2 __% P1 __% P2 __%
********************* *********************
Against itself Against itself
P1 __% P2 __% P1 __% P2 __%
********************* *********************
6x6: ( boards) 6x6: ( boards)
loss (mse) _.___ loss (mse) _.___
Against random Against random
P1 __% P2 __% P1 __% P2 __%
Against greedy Against greedy
P1 __% P2 __% P1 __% P2 __%
********************* *********************
Against itself Against itself
P1 __% P2 __% P1 __% P2 __%
********************* *********************
9/20 update:
How do the nets perform on a *small* amount of 3x3 boards?
(98, not nearly enough to train on)
____________________________________________________________________________________________________________
TEST NO _:
____________________________________________________________________________________________________________
FEATURES: _
ARCHITECTURE: _
BOARD GEN: _
_______________
3x3: (98 boards)
loss (mse) _.____
Against itself
P1 __% P2 __%
Against greedy
P1 __% P2 __%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 29: Convolutional
____________________________________________________________________________________________________________
FEATURES: map of taken
ARCHITECTURE: convolutional
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.3216
Against itself
P1 23% P2 77%
Against greedy
P1 37% P2 56%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 30: Convolutional featured
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0449
Against itself
P1 87% P2 13%
Against greedy
P1 71% P2 36%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 31: Straight-up featured
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0028
Against itself
P1 64% P2 36%
Against greedy
P1 65% P2 40%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 32: Convolutional featured (abstract layer size: n)
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0449
Against itself
P1 87% P2 13%
Against greedy
P1 71% P2 36%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 32: Convolutional featured (abstract layer size: n**2)
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0391
Against itself
P1 28% P2 71%
Against greedy
P1 55% P2 43%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 33: Convolutional featured (abstract layer size: n) (Additional dropout layer)
____________________________________________________________________________________________________________
FEATURES: parity flags for row, col, diag (taken); number of taken, dead, available squares; map of taken squares
ARCHITECTURE: convolutional (map), feeding into flat
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0680
Against itself
P1 95% P2 05%
Against greedy
P1 75% P2 28%
____________________________________________________________________________________________________________
____________________________________________________________________________________________________________
TEST NO 34: Straight-up featured (Additional dropout layer)
____________________________________________________________________________________________________________
FEATURES: tempo flag; number of squares (dead); parity flags for row, col, diag (taken); map of squares (taken)
ARCHITECTURE: flat, through a merged layer
BOARD GEN: init->nbhd->flipped
_______________
3x3: (98 boards)
loss (mse) 0.0513
Against itself
P1 61% P2 39%
Against greedy
P1 61% P2 36%
____________________________________________________________________________________________________________
--------------------------------------------------------------
PRELIMINARY TESTS (basic tree search, learn from training set)
--------------------------------------------------------------
---------------------------- ------------------------------------------------------------
CONV NET CONV + FEATURES + ABSTRACT LAYER (size n) without DROPOUT
---------------------------- ------------------------------------------------------------
3x3: (155 boards) 3x3: (155 boards) <-- about a fifth of the board space
loss (mse) 0.0396 loss (mse) 0.0274
Against itself Against itself
P1 92% P2 08% P1 100% P2 0% <-- the featnet seems to have figured out the strategy
Against featured Against conv
P1 100% P2 1% P1 99% P2 0% <-- but this demonstrates that *both* fail when they aren't the same player
Against random Against random
P1 90% P2 84% P1 94% P2 82%
Against greedy Against greedy
P1 88% P2 20% P1 79% P2 26% <-- shows that *both* fail when they aren't the same player.
4x4: (2378 boards) 4x4: (2378 boards)
loss (mse) 0.0078 loss (mse) 0.0237
Against itself Against itself
P1 92% P2 8% P1 8% P2 92%
Against featured Against conv
P1 18% P2 82% P1 18% P2 82%
Against random Against random
P1 91% P2 96% P1 93% P2 94%
Against greedy Against greedy
P1 49% P2 54% P1 37% P2 55%
5x5: (9958 boards) 5x5: (9958 boards)
loss (mse) 0.0014 loss (mse) _.___
Against itself Against itself
P1 41% P2 59% P1 59% P2 41%
Against featured Against conv
P1 47% P2 49% P1 51% P2 53%
Against random Against random
P1 99% P2 96% P1 99% P2 99%
Against greedy Against greedy
P1 60% P2 39% P1 54% P2 48%
6x6: (79055 boards) 6x6: (79055 boards)
loss (mse) _.___ loss (mse) _.___
Against itself Against itself
P1 69% P2 31% P1 23% P2 77%
Against featured Against conv
P1 33% P2 53% P1 47% P2 67%
Against random Against random
P1 99% P2 98% P1 100% P2 99%
Against greedy Against greedy
P1 55% P2 49% P1 52% P2 47%
Update:
I feel like such an idiot. When I renamed 'greedy' and 'random',
I accidentally labelled the board scores using 'random' instead
of 'greedy'. This resulted in much weaker board scores, which
also explains why my net was performing so badly against the greedy algo.
9/28/2021 UPDATE:
I've been thinking about the game strategy a lot lately. One thing that
Thane Plambeck (the inventor of the game) mentions in his video
http://markhuckvale.com/games/notakto/help.html
is that the winning strategies for 3x3 and 4x4 boards are based on
the player's *response* to their opponent's most recent move.
My feedforward dense network can't learn responses to boards, since
it just reads the static board as it's given. So probably we should use
an LSTM here, or have some other mechanism by which we feed the net the
most recent opponent's move as input.
Post-quals tests on a flat network:
-----------------------
-----------------------
Loss: _.____
Against itself:
P1: _% P2: _%
Against greedy:
P1: _% P2: _%
3x3 boards
-----------------------
1) Just a flat net (size n**2)
-----------------------
Loss: 0.0011
Against itself:
P1: 28% P2: 72%
Against greedy:
P1: 64% P2: 44%
-----------------------
2) Flat net (size n)
-----------------------
Loss: 0.0272