🚀 go-pugleaf

I have spent the last 3 weeks running gnuchess 4.0 pl73 with different
flags through the test suite in the test/CHESSTEST subdirectory.  Four
different versions of gnuchess were tested.  They are

a.  gnuchess 4.0pl73 using NULLMOVE heuristics.  I'll call this NULL.

b.  gnuchess 4.0pl73 using DEEPNULL heuristics.  I'll call this DEEP.
    For this test, I fixed one tiny bug in the DEEPNULL code.  Otherwise
    everything else is left untouched.

c.  gnuchess 4.0pl73 using a slightly modified NULLMOVE heuristics.
    The reason why I'm testing this is because I feel that the present
    null move heuristic is too conservative (it uses ply > 2 and
    depth > 3 before trying the null move).  I changed it to ply > 1
    and depth > 2 and also reduced the material score in the zugzwang
    test.  Overall, this means the null move is tried much more often
    than in the original code.  I'll call this NULL-MODIFIED.

d.  MaxII, the version of gnuchess 4.0 which is running on ICS.  MaxII
    is basically gnuchess 4.0 pl73, but with a few changes in the 
    evaluation routines and search routines.  Some of the more important
    changes include things forward pruning, more bonuses for pawn
    advances, improved endgame play (I think), faster mate searches and
    so on.  MaxII also uses the NULL-MODIFIED heuristics which is mentioned
    in (c) above.  This test was carried out as an afterthought as I wanted
    to see whether my tinkering has produced any improvement.

The hardware for this test is one R4400 (150 Mhz) processor of an Onyx
system.  This chip is rated at 88 SPECint92.  Depending on which null move
heuristic is being tested, the average speed is about 20K-25K nodes/sec.
The compiler used is GCC 2.5.8.

Here are the results of the four tests.  To briefly summarize, the worst
performance (2864) was posted by the NULL version which is what I expected.
The NULL-MODIFIED version did much better to produced a score of 2950.
However, in 3 of the test suites (jenoban, larsen3 and speelman), it 
performed worse than the NULL version.  All these three test suites are
endgame tests.  My conclusion here was that the material value used in
the zugzwang test during the null move was too small (I was using a value
of the Queen).  The way to test this guess would be to increase the
zugzwang material value and rerun the NULL-MODIFIED version on the 3 test
suites.

The DEEP version did better than the NULL and NULL-MODIFIED, scoring 2978,
which is slightly better than the NULL-MODIFIED.  Again, there are some
test suites in which it performed worst than the NULL-MODIFIED.  I didn't
try to figure this one out.  

The MaxII version did much better than all the 3 above versions, scoring
3163.  Only in 2 testsuites did it did slightly worse, and that is only 
by 1 position, which is not significant.   The increase scores in the
tactical suites like gelfer and hardmid is significant enough to make me
believe MaxII is tactically stronger.


Kong Sian



1.  Result for NULL version

  test         		right   wrong  total    percent
CHESSTEST/bellin	6	27	33	18
CHESSTEST/ece3		848	949	1797	47
CHESSTEST/finecomb	12	38	50	24
CHESSTEST/gelfer	84	243	327	26
CHESSTEST/hardmid	411	467	878	47
CHESSTEST/hortjan	59	74	133	44
CHESSTEST/jenoban	125	77	202	62
CHESSTEST/kcmdm		41	105	146	28
CHESSTEST/kotov		40	75	115	35
CHESSTEST/larsen1	91	9	100	91
CHESSTEST/larsen2	11	37	48	23
CHESSTEST/larsen3	31	32	63	49
CHESSTEST/openings	34	76	110	31
CHESSTEST/positional	50	104	154	32
CHESSTEST/reinfeld	688	311	999	69
CHESSTEST/sherev	10	50	60	17
CHESSTEST/speelman	65	102	167	39
CHESSTEST/winatchess	258	42	300	86

total    		2864	2818	5682	50



2.  Result for NULL-MODIFIED

  test         		right   wrong  total    percent
CHESSTEST/bellin	6	27	33	18
CHESSTEST/ece3		883	914	1797	49
CHESSTEST/finecomb	12	38	50	24
CHESSTEST/gelfer	86	241	327	26
CHESSTEST/hardmid	433	445	878	49
CHESSTEST/hortjan	65	68	133	49
CHESSTEST/jenoban	114	88	202	56
CHESSTEST/kcmdm		46	100	146	32
CHESSTEST/kotov		43	72	115	37
CHESSTEST/larsen1	91	9	100	91
CHESSTEST/larsen2	12	36	48	25
CHESSTEST/larsen3	30	33	63	48
CHESSTEST/openings	34	76	110	31
CHESSTEST/positional	54	100	154	35
CHESSTEST/reinfeld	704	295	999	70
CHESSTEST/sherev	10	50	60	17
CHESSTEST/speelman	60	106	166	36
CHESSTEST/winatchess	267	33	300	89

total    		2950	2731	5681	52


3.  Results for DEEP version.

  test         		right   wrong  total    percent
CHESSTEST/bellin	8	25	33	24
CHESSTEST/ece3		898	899	1797	50
CHESSTEST/finecomb	13	37	50	26
CHESSTEST/gelfer	86	241	327	26
CHESSTEST/hardmid	436	442	878	50
CHESSTEST/hortjan	65	68	133	49
CHESSTEST/jenoban	126	76	202	62
CHESSTEST/kcmdm		40	106	146	27
CHESSTEST/kotov		42	73	115	37
CHESSTEST/larsen1	91	9	100	91
CHESSTEST/larsen2	14	34	48	29
CHESSTEST/larsen3	28	35	63	44
CHESSTEST/openings	34	76	110	31
CHESSTEST/positional	50	104	154	32
CHESSTEST/reinfeld	703	296	999	70
CHESSTEST/sherev	14	46	60	23
CHESSTEST/speelman	65	102	167	39
CHESSTEST/winatchess	265	35	300	88

total    		2978	2704	5682	52



4.  Results for the MaxII version

  	test         right   wrong  total    percent
CHESSTEST/bellin	9	24	33	27
CHESSTEST/ece3		957	840	1797	53
CHESSTEST/finecomb	13	37	50	26
CHESSTEST/gelfer	101	226	327	31
CHESSTEST/hardmid	469	409	878	53
CHESSTEST/hortjan	73	60	133	55
CHESSTEST/jenoban	132	70	202	65
CHESSTEST/kcmdm		45	101	146	31
CHESSTEST/kotov		44	71	115	38
CHESSTEST/larsen1	91	9	100	91
CHESSTEST/larsen2	15	33	48	31
CHESSTEST/larsen3	29	34	63	46
CHESSTEST/openings	51	59	110	46
CHESSTEST/positional	58	95	153	38
CHESSTEST/reinfeld	709	290	999	71
CHESSTEST/sherev	15	45	60	25
CHESSTEST/speelman	80	87	167	48
CHESSTEST/winatchess	272	28	300	91

total    		3163	2518	5681	56