The Distributed Chess Project

Creating Chess-Playing Artificial Neural Networks with Distributed Evolutionary Algorithms

Since 06/23/2002 a cumulated computing time of 996 days 5 hours 21 min 24 sec has been contributed.

90 participants have returned at least one candidate neural network. The results of the best ones are ranked according to test performance, corresponding training performance and contributed computing time in the following table:

Rank	User-ID	Best Performance on Test Set (%)	Best Performance on Training Set (%)	Cumulated Computing Time (d:h:m:s)
1	esa.elovaara	78	88	83:2:20:38
2	jason	78	88	7:3:22:3
3	peter.k.campbell	76	94	57:3:32:24
4	henkf	76	94	8:11:10:50
5	midmalex	76	92	59:6:56:35
6	regloh.lliw	76	92	3:18:20:48
7	ssieh	76	92	1:2:9:47
8	freemac_99	76	90	50:10:1:26
9	scotte	76	90	9:23:21:39
10	pt.schwarz	76	90	7:23:16:44
11	thornton	76	90	5:13:7:57
12	stefan.bengtsson	76	88	2:10:55:26
13	lithron	76	88	1:22:39:52
14	W	76	86	14:11:55:25
15	gipe	76	86	2:22:34:39
16	sr	76	82	9:20:9:6
17	amiguel	74	94	11:22:16:33
18	m_diosi	74	92	32:10:9:6
19	tran_tang	74	92	21:12:11:31
20	ad_lord\hotmail.com	74	92	17:12:17:24
21	atalmadge	74	92	13:4:22:25
22	john.obrien	74	92	3:15:59:17
23	NeuroBernd	74	92	2:3:50:41
24	u37206657	74	90	32:12:47:4
25	simon	74	90	30:7:54:55
26	vishvananda	74	90	29:6:21:43
27	njenson	74	90	17:7:5:26
28	richbcla	74	90	12:11:0:29
29	david7091	74	90	9:8:34:20
30	5xd7g251lq2vs001	74	90	7:19:49:30
31	gillrich	74	90	3:13:12:34
32	ddcc	74	90	2:21:10:12
33	shop	74	90	2:4:18:42
34	qwerty4	74	88	15:6:31:45
35	jarekfil	74	88	3:21:51:7
36	regloh.lliw	74	88	2:12:22:10
37	mitchellonline	74	86	8:0:19:27
38	jean_efpraxiadis	74	84	16:15:9:32
39	rich_luce	74	84	8:20:6:46
40	arnowa	72	94	6:14:56:2
41	ravik	72	94	3:12:37:27
42	thunder124	72	92	12:2:33:35
43	ssmythe	72	92	8:12:34:10
44	zanglerska	72	92	3:21:14:39
45	santzi	72	92	2:13:44:23
46	toddkloos	72	92	1:18:49:49
47	Sebastian	72	92	1:9:25:17
48	llclarisse	72	90	31:2:34:44
49	ic	72	90	26:2:31:58
50	organizer	72	90	23:20:15:24
51	mutefly	72	90	15:9:50:32
52	wingy	72	90	13:1:48:49
53	joachim	72	90	10:17:26:20
54	idiott47	72	90	8:12:17:43
55	kannan	72	90	7:10:12:14
56	alasdair	72	90	6:1:17:58
57	kennynthebest	72	90	6:0:19:37
58	jopicard	72	90	4:20:22:12
59	azrael	72	90	4:18:7:10
60	fronte	72	90	3:7:31:4
61	doo	72	90	2:16:45:29
62	frank	72	90	2:4:59:41
63	sliver	72	90	1:21:28:23
64	vince.mele	72	90	1:18:2:21
65	aaron	72	90	0:23:47:58
66	holle	72	88	17:9:56:33
67	jwa	72	88	12:15:43:11
68	mentis	72	86	20:1:9:18
69	a.wagner1	72	86	5:7:23:57
70	petera	72	86	0:5:7:53
71	afmomania	72	86	0:3:27:12
72	mhond99	72	84	6:8:55:35
73	turrinipa	72	82	10:0:29:33
74	layton97	72	80	3:17:16:44
75	amanwithattitude	70	94	1:23:3:7
76	rdmltrs	70	92	9:20:46:55
77	arno	70	92	3:10:10:46
78	lshome	70	92	3:3:59:49
79	davidroel	70	90	5:19:51:36
80	NNBernd	70	90	3:7:49:43
81	paul.abrams	70	90	1:20:1:41
82	joe	70	90	1:13:57:13
83	stjernbecker	70	90	0:22:20:39
84	berndt	70	86	2:5:4:32
85	chess	68	94	5:12:34:34
86	agarcia	68	92	14:14:36:18
87	mljwing	68	92	12:20:16:30
88	Barrios_Blue	68	92	6:1:41:49
89	ttf11	68	90	2:12:35:23
90	pouilley	68	88	2:15:47:56

What I call the 'test set' consists of 50 chess problems with known best continuations. This set must be viewed in contrast to the 'training set' consisting of 50 different problems. The test problems are presented to the net after the training is finished in order to test how well a neural net generalizes from the training problems. This is most important, since a neural net that does well on the training set but not on an independent test set isn't worth its money.

And this is the good news: There definitely is generalization to some degree in the results so far. During training, the neural nets do in fact capture some positional patterns, that enable them to find the correct continuations most of the time when confronted with the test problems.

Numerically, the results in the table are encouraging and rather impressive, because a neural net with e.g. 75% test performance picks the only correct move from a variety of 30 to 50 possible legal moves for 3 out of 4 problems it sees for the very first time. From the pessimist point of view, however, it still fails on 1 out of 4 problems. Moreover, since the typical training performance is higher, we don't have perfect generalization. Therefore it must be the goal to find neural nets with test performance as high as the training performance.

What you as active project participants can and should do at this point is fiddle around with the free parameters of both neural networks and genetic algorithm in order to find a configuration that maximizes the test performance. Any questions? Please check the documentation first, but you may contact me directly as well via Ralf.Seliger@t-online.de