36:
512:
The neural network used for the original 2018 computer shogi implementation consists of four weight layers: W1 (16-bit integers) and W2, W3 and W4 (8-bit). It has 4 fully-connected layers,
430:
122:
652:
475:
tree. While being slower than handcrafted evaluation functions, NNUE does not suffer from the 'blindness beyond the current move' problem.
479:
81:
423:
45:
97:
761:
693:
416:
671:"Efficiently Updatable Neural-Network-based Evaluation Function for computer Shogi (Unofficial English Translation)"
519:
W1 encoded the king's position and therefore this layer needed only to be re-evaluated once the king moved. It used
708:
217:
766:
87:
756:
513:
471:, or variants thereof like the king-piece-square table. NNUE is used primarily for the leaf nodes of the
112:
599:
212:
162:
572:
232:
506:
776:
771:
520:
502:
172:
551:
545:
528:
472:
347:
322:
282:
107:
192:
127:
392:
8:
540:
464:
372:
117:
77:
70:
670:
468:
157:
102:
487:
387:
167:
624:
653:"Efficiently Updatable Neural-Network-based Evaluation Function for computer Shogi"
495:
337:
297:
516:
activation functions, and outputs a single number, being the score of the board.
486:
in 2018. On 6 August 2020, NNUE was for the first time ported to a chess engine,
312:
140:
491:
483:
460:
332:
317:
185:
144:
57:
27:
750:
402:
377:
362:
292:
227:
357:
250:
197:
272:
262:
731:
741:
342:
277:
207:
490:
12. Since 2021, many of the top rated classical chess engines such as
397:
382:
327:
302:
287:
257:
35:
202:
737:
152:
713:
678:
577:
352:
307:
267:
237:
222:
367:
524:
52:
548:- The chapter about NNUE features a visualization of NNUE.
452:
494:
have an NNUE implementation to remain competitive (with
706:
748:
570:
424:
738:NNUE evaluation functions for computer shogi
709:"official-stockfish / Stockfish, NNUE merge"
431:
417:
527:(SIMD) techniques along with appropriate
16:Neural network based evaluation function
668:
650:
749:
646:
644:
93:Efficiently updatable neural networks
22:This article is part of the series on
707:Joost VandeVondele (July 25, 2020).
662:
445:efficiently updatable neural network
641:
625:"Stockfish - Chessprogramming wiki"
13:
14:
788:
725:
571:Gary Linscott (April 30, 2021).
525:single instruction multiple data
98:Handcrafted evaluation functions
34:
734:on the Chess Programming Wiki.
700:
686:
617:
592:
564:
1:
694:"Introducing NNUE Evaluation"
557:
505:without a requirement for a
7:
534:
113:Stochastic gradient descent
10:
793:
762:Artificial neural networks
669:Yu Nasu (April 28, 2018).
651:Yu Nasu (April 28, 2018).
163:Principal variation search
501:NNUE runs efficiently on
498:as a notable exception).
451:, a Japanese wordplay on
629:www.chessprogramming.org
507:graphics processing unit
503:central processing units
455:, sometimes stylised as
521:incremental computation
173:Monte Carlo tree search
552:List of chess software
546:Stockfish chess engine
529:intrinsic instructions
283:Dragon by Komodo Chess
108:Reinforcement learning
478:NNUE was invented by
128:Unsupervised learning
46:Board representations
78:Deep neural networks
71:Evaluation functions
767:Japanese inventions
541:elmo (shogi engine)
469:piece-square tables
465:evaluation function
118:Supervised learning
103:Piece-square tables
757:Evaluation methods
482:and introduced to
158:Alpha-beta pruning
467:whose inputs are
441:
440:
168:Quiescence search
147:search algorithms
28:Chess programming
784:
719:
718:
704:
698:
697:
696:. 6 August 2020.
690:
684:
683:
675:
666:
660:
659:
657:
648:
639:
638:
636:
635:
621:
615:
614:
612:
610:
596:
590:
589:
587:
585:
568:
496:Leela Chess Zero
433:
426:
419:
338:Leela Chess Zero
38:
19:
18:
792:
791:
787:
786:
785:
783:
782:
781:
747:
746:
728:
723:
722:
705:
701:
692:
691:
687:
673:
667:
663:
655:
649:
642:
633:
631:
623:
622:
618:
608:
606:
598:
597:
593:
583:
581:
569:
565:
560:
537:
437:
408:
407:
253:
243:
242:
188:
186:Chess computers
178:
177:
148:
133:
132:
73:
63:
62:
48:
17:
12:
11:
5:
790:
780:
779:
777:Computer chess
774:
772:Computer shogi
769:
764:
759:
745:
744:
735:
727:
726:External links
724:
721:
720:
699:
685:
661:
658:(in Japanese).
640:
616:
604:Stockfish Blog
600:"Stockfish 12"
591:
562:
561:
559:
556:
555:
554:
549:
543:
536:
533:
484:computer shogi
461:neural network
439:
438:
436:
435:
428:
421:
413:
410:
409:
406:
405:
400:
395:
390:
385:
380:
375:
370:
365:
360:
355:
350:
345:
340:
335:
330:
325:
320:
315:
310:
305:
300:
295:
290:
285:
280:
275:
270:
265:
260:
254:
249:
248:
245:
244:
241:
240:
235:
230:
225:
220:
215:
210:
205:
200:
195:
189:
184:
183:
180:
179:
176:
175:
170:
165:
160:
155:
149:
139:
138:
135:
134:
131:
130:
125:
120:
115:
110:
105:
100:
95:
90:
85:
74:
69:
68:
65:
64:
61:
60:
55:
49:
44:
43:
40:
39:
31:
30:
24:
23:
15:
9:
6:
4:
3:
2:
789:
778:
775:
773:
770:
768:
765:
763:
760:
758:
755:
754:
752:
743:
739:
736:
733:
730:
729:
716:
715:
710:
703:
695:
689:
681:
680:
672:
665:
654:
647:
645:
630:
626:
620:
605:
601:
595:
580:
579:
574:
567:
563:
553:
550:
547:
544:
542:
539:
538:
532:
530:
526:
522:
517:
515:
510:
508:
504:
499:
497:
493:
492:Komodo Dragon
489:
485:
481:
476:
474:
470:
466:
462:
458:
454:
450:
446:
434:
429:
427:
422:
420:
415:
414:
412:
411:
404:
401:
399:
396:
394:
391:
389:
386:
384:
381:
379:
376:
374:
371:
369:
366:
364:
361:
359:
356:
354:
351:
349:
346:
344:
341:
339:
336:
334:
331:
329:
326:
324:
321:
319:
316:
314:
311:
309:
306:
304:
301:
299:
296:
294:
291:
289:
286:
284:
281:
279:
276:
274:
271:
269:
266:
264:
261:
259:
256:
255:
252:
251:Chess engines
247:
246:
239:
236:
234:
231:
229:
226:
224:
221:
219:
216:
214:
211:
209:
206:
204:
201:
199:
196:
194:
191:
190:
187:
182:
181:
174:
171:
169:
166:
164:
161:
159:
156:
154:
151:
150:
146:
142:
137:
136:
129:
126:
124:
121:
119:
116:
114:
111:
109:
106:
104:
101:
99:
96:
94:
91:
89:
86:
83:
79:
76:
75:
72:
67:
66:
59:
56:
54:
51:
50:
47:
42:
41:
37:
33:
32:
29:
26:
25:
21:
20:
712:
702:
688:
677:
664:
632:. Retrieved
628:
619:
607:. Retrieved
603:
594:
584:December 12,
582:. Retrieved
576:
566:
518:
511:
500:
477:
456:
448:
444:
442:
218:Deep Thought
198:ChessMachine
123:Texel tuning
92:
82:Transformers
273:CuckooChess
263:Chess Tiger
751:Categories
742:github.com
634:2020-08-18
609:19 October
558:References
473:alpha–beta
343:MChess Pro
278:Deep Fritz
208:Cray Blitz
488:Stockfish
398:Turochamp
388:Stockfish
383:SmarThink
328:KnightCap
303:GNU Chess
288:Fairy-Max
258:AlphaZero
213:Deep Blue
88:Attention
58:Bitboards
535:See also
509:(GPU).
373:Shredder
233:Mephisto
203:ChipTest
480:Yu Nasu
463:-based
459:) is a
348:Mittens
313:Houdini
153:Minimax
714:GitHub
679:GitHub
578:GitHub
573:"NNUE"
353:MuZero
333:Komodo
323:Junior
318:Ikarus
308:HIARCS
268:Crafty
238:Saitek
223:HiTech
674:(PDF)
656:(PDF)
403:Zappa
393:Torch
378:Sjeng
368:Rybka
363:REBEL
298:Fruit
293:Fritz
228:Hydra
193:Belle
141:Graph
732:NNUE
611:2020
586:2020
523:and
514:ReLU
457:ƎUИИ
449:NNUE
358:Naum
145:tree
143:and
53:0x88
740:on
453:Nue
443:An
753::
711:.
676:.
643:^
627:.
602:.
575:.
531:.
717:.
682:.
637:.
613:.
588:.
447:(
432:e
425:t
418:v
84:)
80:(
Text is available under the Creative Commons Attribution-ShareAlike License. Additional terms may apply.