Is there a program that can play better simply by playing against itself and learning from the games?
The learning that I think about is learning to have better weights.
It is clear that the weights of the program are not optimal.
A program can try to change weights and test the value of the change in games against itself.
It is possible to test evaluation changes by games of the program against itself at small fixed depth.
depth1,depth 2,depth 3 and depth 4 with 200,000 games in every match.
The point is not to find a proved rating improvement based on small depth but to find a reason to believe that there is a rating improvement that you need to test later.
probably you do not get rating improvement if you find something like
100700/200000 at depth 1
100400/200000 at depth 2
100200/200000 at depth 3
100050/200000 at depth 4
On the other hand you can be optimistic about rating improvement if you see something like
99300/200000 at depth 1
99600/200000 at depth 2
99800/200000 at depth 3
99950/200000 at depth 4
I believe that it may be possible to make stockfish at least 10 elo better based on some automatic learning from games against itself when you change weights when the first games are at very small depths but of course somebody need to write the relevant code and later give it a month of computer time.
What is your opinion?
learning question about programs
Moderators: hgm, Rebel, chrisw
-
- Posts: 10413
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
-
- Posts: 4840
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: learning question about programs
By using clop and cutechess-cli, what you said can almost be done by all engines. Say you want to optimize a certain param value, allowing clop to test at range 0 to 100, I have a tool to parse the clop data result and get the result stat of every parameter value tested, see sample below. At one time I was optimizing my lmr factor, I use clop at param range 10 to 200, and allow it to play my default at lmr factor of 20.
So what clop does is play param 10 vs 20, param 11 vs 20, and so on. I use TC 40 moves/10s repeating. After a total of 5067 games, param value 64 got a good performance (after 36 games of this particular param value). Top performers are listed at the bottom.
I may try your system using fixed depth.
So what clop does is play param 10 vs 20, param 11 vs 20, and so on. I use TC 40 moves/10s repeating. After a total of 5067 games, param value 64 got a good performance (after 36 games of this particular param value). Top performers are listed at the bottom.
I may try your system using fixed depth.
Code: Select all
CLOP Data Reader v3.0
Mar 11 2014, 22:27:47
Number of parameters: 1
First parameter: LmrFactor
Param1: Min 10, Max 200
Total games: 5067
Param1 W / L / D NetW Games Score LOS
10 9 / 6 / 9, +3 24 56.25% 77.28%
11 11 / 5 / 24, +6 40 57.50% 92.83%
12 17 / 12 / 21, +5 50 55.00% 81.92%
13 10 / 4 / 22, +6 36 58.33% 94.08%
14 4 / 9 / 15, -5 28 41.07% 8.98%
15 13 / 10 / 21, +3 44 53.41% 72.94%
16 16 / 6 / 18, +10 40 62.50% 98.27%
17 12 / 6 / 16, +6 34 58.82% 91.65%
18 9 / 11 / 22, -2 42 47.62% 33.18%
19 17 / 6 / 17, +11 40 63.75% 98.87%
20 10 / 8 / 18, +2 36 52.78% 67.62%
21 11 / 6 / 17, +5 34 57.35% 88.11%
22 2 / 3 / 5, -1 10 45.00% 34.38%
23 17 / 7 / 12, +10 36 63.89% 97.84%
24 9 / 9 / 22, +0 40 50.00% 50.00%
25 14 / 8 / 28, +6 50 56.00% 89.50%
26 8 / 5 / 13, +3 26 55.77% 78.80%
27 9 / 4 / 9, +5 22 61.36% 91.02%
28 12 / 9 / 21, +3 42 53.57% 73.83%
29 14 / 7 / 15, +7 36 59.72% 93.31%
30 14 / 4 / 20, +10 38 63.16% 99.04%
31 7 / 7 / 24, +0 38 50.00% 50.00%
32 19 / 6 / 19, +13 44 64.77% 99.53%
33 6 / 6 / 8, +0 20 50.00% 50.00%
34 13 / 11 / 20, +2 44 52.27% 65.50%
35 4 / 6 / 10, -2 20 45.00% 27.44%
36 3 / 1 / 6, +2 10 60.00% 81.25%
37 9 / 5 / 22, +4 36 55.56% 84.91%
38 7 / 8 / 11, -1 26 48.08% 40.18%
39 7 / 4 / 11, +3 22 56.82% 80.62%
40 10 / 5 / 15, +5 30 58.33% 89.49%
41 5 / 4 / 11, +1 20 52.50% 62.30%
42 5 / 7 / 6, -2 18 44.44% 29.05%
43 11 / 8 / 17, +3 36 54.17% 74.83%
44 11 / 12 / 17, -1 40 48.75% 41.94%
45 15 / 11 / 22, +4 48 54.17% 77.90%
46 4 / 4 / 12, +0 20 50.00% 50.00%
47 9 / 9 / 20, +0 38 50.00% 50.00%
48 2 / 7 / 11, -5 20 37.50% 5.47%
49 6 / 7 / 17, -1 30 48.33% 39.53%
50 4 / 9 / 19, -5 32 42.19% 8.98%
51 14 / 6 / 20, +8 40 60.00% 96.08%
52 16 / 9 / 19, +7 44 57.95% 91.57%
53 10 / 10 / 26, +0 46 50.00% 50.00%
54 11 / 8 / 19, +3 38 53.95% 74.83%
55 3 / 5 / 4, -2 12 41.67% 25.39%
56 13 / 10 / 19, +3 42 53.57% 72.94%
57 13 / 10 / 21, +3 44 53.41% 72.94%
58 11 / 14 / 21, -3 46 46.74% 27.86%
59 11 / 4 / 23, +7 38 59.21% 96.16%
60 9 / 7 / 12, +2 28 53.57% 68.55%
61 9 / 6 / 25, +3 40 53.75% 77.28%
62 6 / 3 / 17, +3 26 55.77% 82.81%
63 7 / 4 / 13, +3 24 56.25% 80.62%
64 18 / 4 / 14, +14 36 69.44% 99.87%
65 8 / 5 / 9, +3 22 56.82% 78.80%
66 6 / 7 / 15, -1 28 48.21% 39.53%
67 8 / 8 / 18, +0 34 50.00% 50.00%
68 4 / 2 / 10, +2 16 56.25% 77.34%
69 6 / 5 / 11, +1 22 52.27% 61.28%
70 12 / 5 / 13, +7 30 61.67% 95.19%
71 8 / 5 / 9, +3 22 56.82% 78.80%
72 5 / 1 / 4, +4 10 70.00% 93.75%
73 7 / 2 / 13, +5 22 61.36% 94.53%
74 2 / 2 / 6, +0 10 50.00% 50.00%
75 6 / 5 / 11, +1 22 52.27% 61.28%
76 7 / 4 / 9, +3 20 57.50% 80.62%
77 14 / 6 / 22, +8 42 59.52% 96.08%
78 6 / 6 / 12, +0 24 50.00% 50.00%
79 5 / 4 / 7, +1 16 53.13% 62.30%
80 11 / 8 / 15, +3 34 54.41% 74.83%
81 9 / 5 / 16, +4 30 56.67% 84.91%
82 4 / 4 / 8, +0 16 50.00% 50.00%
83 2 / 5 / 19, -3 26 44.23% 14.45%
84 3 / 7 / 18, -4 28 42.86% 11.33%
85 1 / 3 / 12, -2 16 43.75% 18.75%
86 5 / 8 / 13, -3 26 44.23% 21.20%
87 5 / 5 / 10, +0 20 50.00% 50.00%
88 10 / 6 / 14, +4 30 56.67% 83.38%
89 8 / 7 / 17, +1 32 51.56% 59.82%
90 6 / 8 / 6, -2 20 45.00% 30.36%
91 10 / 7 / 15, +3 32 54.69% 75.97%
92 7 / 7 / 24, +0 38 50.00% 50.00%
93 7 / 3 / 12, +4 22 59.09% 88.67%
94 5 / 2 / 7, +3 14 60.71% 85.55%
95 1 / 0 / 5, +1 6 58.33% 75.00%
96 4 / 4 / 6, +0 14 50.00% 50.00%
97 4 / 1 / 13, +3 18 58.33% 89.06%
98 6 / 2 / 16, +4 24 58.33% 91.02%
99 6 / 3 / 5, +3 14 60.71% 82.81%
100 7 / 8 / 11, -1 26 48.08% 40.18%
101 8 / 7 / 17, +1 32 51.56% 59.82%
102 6 / 2 / 10, +4 18 61.11% 91.02%
103 2 / 3 / 7, -1 12 45.83% 34.38%
104 7 / 5 / 16, +2 28 53.57% 70.95%
105 5 / 8 / 7, -3 20 42.50% 21.20%
106 5 / 5 / 10, +0 20 50.00% 50.00%
107 5 / 5 / 12, +0 22 50.00% 50.00%
108 8 / 3 / 9, +5 20 62.50% 92.70%
109 4 / 3 / 5, +1 12 54.17% 63.67%
110 6 / 4 / 18, +2 28 53.57% 72.56%
111 3 / 1 / 4, +2 8 62.50% 81.25%
112 3 / 9 / 8, -6 20 35.00% 4.61%
113 7 / 4 / 15, +3 26 55.77% 80.62%
114 1 / 2 / 1, -1 4 37.50% 31.25%
115 8 / 5 / 9, +3 22 56.82% 78.80%
116 6 / 5 / 9, +1 20 52.50% 61.28%
117 1 / 0 / 1, +1 2 75.00% 75.00%
118 8 / 6 / 20, +2 34 52.94% 69.64%
119 9 / 5 / 8, +4 22 59.09% 84.91%
120 7 / 5 / 20, +2 32 53.13% 70.95%
121 3 / 9 / 4, -6 16 31.25% 4.61%
122 4 / 3 / 3, +1 10 55.00% 63.67%
123 2 / 8 / 10, -6 20 35.00% 3.27%
124 8 / 9 / 15, -1 32 48.44% 40.73%
125 8 / 10 / 16, -2 34 47.06% 32.38%
126 6 / 4 / 16, +2 26 53.85% 72.56%
127 4 / 4 / 6, +0 14 50.00% 50.00%
128 11 / 6 / 13, +5 30 58.33% 88.11%
129 6 / 4 / 18, +2 28 53.57% 72.56%
130 8 / 4 / 22, +4 34 55.88% 86.66%
131 12 / 16 / 38, -4 66 46.97% 22.91%
132 4 / 5 / 11, -1 20 47.50% 37.70%
133 5 / 6 / 7, -1 18 47.22% 38.72%
134 8 / 8 / 22, +0 38 50.00% 50.00%
135 3 / 3 / 8, +0 14 50.00% 50.00%
136 5 / 6 / 9, -1 20 47.50% 38.72%
137 9 / 2 / 15, +7 26 63.46% 98.07%
138 5 / 4 / 7, +1 16 53.13% 62.30%
139 6 / 3 / 9, +3 18 58.33% 82.81%
140 10 / 7 / 21, +3 38 53.95% 75.97%
141 8 / 3 / 11, +5 22 61.36% 92.70%
142 11 / 9 / 18, +2 38 52.63% 66.82%
143 5 / 4 / 9, +1 18 52.78% 62.30%
144 3 / 1 / 6, +2 10 60.00% 81.25%
145 6 / 3 / 5, +3 14 60.71% 82.81%
146 7 / 1 / 8, +6 16 68.75% 98.05%
147 7 / 7 / 14, +0 28 50.00% 50.00%
148 3 / 5 / 10, -2 18 44.44% 25.39%
149 5 / 9 / 10, -4 24 41.67% 15.09%
150 6 / 5 / 25, +1 36 51.39% 61.28%
151 9 / 4 / 5, +5 18 63.89% 91.02%
152 5 / 0 / 7, +5 12 70.83% 98.44%
153 3 / 3 / 4, +0 10 50.00% 50.00%
154 6 / 7 / 13, -1 26 48.08% 39.53%
155 11 / 4 / 9, +7 24 64.58% 96.16%
156 5 / 3 / 12, +2 20 55.00% 74.61%
157 6 / 7 / 9, -1 22 47.73% 39.53%
158 2 / 3 / 5, -1 10 45.00% 34.38%
159 11 / 9 / 14, +2 34 52.94% 66.82%
160 8 / 4 / 12, +4 24 58.33% 86.66%
161 7 / 9 / 10, -2 26 46.15% 31.45%
162 6 / 14 / 16, -8 36 38.89% 3.92%
163 3 / 4 / 7, -1 14 46.43% 36.33%
164 5 / 7 / 10, -2 22 45.45% 29.05%
165 8 / 11 / 13, -3 32 45.31% 25.17%
166 11 / 8 / 17, +3 36 54.17% 74.83%
167 11 / 12 / 29, -1 52 49.04% 41.94%
168 3 / 4 / 9, -1 16 46.88% 36.33%
169 6 / 4 / 14, +2 24 54.17% 72.56%
170 4 / 7 / 9, -3 20 42.50% 19.38%
171 5 / 3 / 12, +2 20 55.00% 74.61%
172 11 / 12 / 21, -1 44 48.86% 41.94%
173 8 / 2 / 12, +6 22 63.64% 96.73%
174 11 / 5 / 14, +6 30 60.00% 92.83%
175 9 / 10 / 17, -1 36 48.61% 41.19%
176 11 / 5 / 14, +6 30 60.00% 92.83%
177 7 / 8 / 21, -1 36 48.61% 40.18%
178 1 / 6 / 11, -5 18 36.11% 3.52%
179 8 / 5 / 15, +3 28 55.36% 78.80%
180 6 / 4 / 2, +2 12 58.33% 72.56%
181 11 / 7 / 10, +4 28 57.14% 82.04%
182 4 / 6 / 16, -2 26 46.15% 27.44%
183 9 / 4 / 21, +5 34 57.35% 91.02%
184 4 / 1 / 9, +3 14 60.71% 89.06%
185 5 / 13 / 12, -8 30 36.67% 3.18%
186 13 / 3 / 12, +10 28 67.86% 99.36%
187 13 / 9 / 24, +4 46 54.35% 79.76%
188 8 / 4 / 10, +4 22 59.09% 86.66%
189 10 / 7 / 11, +3 28 55.36% 75.97%
190 8 / 3 / 7, +5 18 63.89% 92.70%
191 11 / 2 / 23, +9 36 62.50% 99.35%
192 7 / 6 / 11, +1 24 52.08% 60.47%
193 3 / 4 / 9, -1 16 46.88% 36.33%
194 6 / 2 / 8, +4 16 62.50% 91.02%
195 4 / 4 / 14, +0 22 50.00% 50.00%
196 7 / 4 / 7, +3 18 58.33% 80.62%
197 5 / 8 / 15, -3 28 44.64% 21.20%
198 4 / 2 / 6, +2 12 58.33% 77.34%
199 14 / 9 / 15, +5 38 56.58% 84.63%
200 3 / 3 / 8, +0 14 50.00% 50.00%
Top Parameters: By LOS
[1] par1 64, score 69.44%, LOS 99.870%, Games 36, NetWins +14
[2] par1 32, score 64.77%, LOS 99.532%, Games 44, NetWins +13
[3] par1 186, score 67.86%, LOS 99.364%, Games 28, NetWins +10
[4] par1 191, score 62.50%, LOS 99.353%, Games 36, NetWins +9
[5] par1 30, score 63.16%, LOS 99.039%, Games 38, NetWins +10
-
- Posts: 7221
- Joined: Mon May 27, 2013 10:31 am
Re: learning question about programs
I thought about replacing evaluation by a neural network. But learning is always terrible slow. And there are perhaps too many weights in the network. Perhaps three layers are enough but you have 64 inputs.
-
- Posts: 38
- Joined: Tue Jul 01, 2008 9:36 pm
Re: learning question about programs
Not by playing against itself, but Bebe did some successful learning. There is a good description in Computers, Chess and Cognition by the Scherzers.Uri Blass wrote:Is there a program that can play better simply by playing against itself and learning from the games?
-
- Posts: 27869
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: learning question about programs
I do this all the time. It is how I get all my piece values. Just not automatically, but it could easily have been automated. When you want to know how much a certain position characteristic (e.g. posession of a piece of material X vs posession of material Y) is worth, I take a collection of positions that only differ in that characteristic, and let the engine play those against itself. Then I know from the result how much the characteristic was worth (in cP), as the Pawn-odds score is known. If that is very different from what the engine assumed, I tune its eval term according to the result, and repeat the tests. (Never met a case where this altered the result, btw. But that might be because the first guess I use is never completely crazy; I am pretty sure that when you would set the Q value to 50 cP you would need more than one iteration to obtain self-consistency.)Uri Blass wrote:Is there a program that can play better simply by playing against itself and learning from the games?
That way a couple of hundred games is usually enough to get an accurate tuning of the major terms, to an accuracy of ~25 cP.
-
- Posts: 303
- Joined: Sat Apr 28, 2012 6:18 pm
- Location: Austin, TX
Re: learning question about programs
Ferdinand,Ferdy wrote:By using clop and cutechess-cli, what you said can almost be done by all engines. Say you want to optimize a certain param value, allowing clop to test at range 0 to 100, I have a tool to parse the clop data result and get the result stat of every parameter value tested, see sample below. At one time I was optimizing my lmr factor, I use clop at param range 10 to 200, and allow it to play my default at lmr factor of 20.
So what clop does is play param 10 vs 20, param 11 vs 20, and so on. I use TC 40 moves/10s repeating. After a total of 5067 games, param value 64 got a good performance (after 36 games of this particular param value). Top performers are listed at the bottom.
I may try your system using fixed depth.Code: Select all
CLOP Data Reader v3.0 Mar 11 2014, 22:27:47 Number of parameters: 1 First parameter: LmrFactor Param1: Min 10, Max 200 Total games: 5067 Param1 W / L / D NetW Games Score LOS 10 9 / 6 / 9, +3 24 56.25% 77.28% 11 11 / 5 / 24, +6 40 57.50% 92.83% 12 17 / 12 / 21, +5 50 55.00% 81.92% 13 10 / 4 / 22, +6 36 58.33% 94.08% 14 4 / 9 / 15, -5 28 41.07% 8.98% 15 13 / 10 / 21, +3 44 53.41% 72.94% 16 16 / 6 / 18, +10 40 62.50% 98.27% 17 12 / 6 / 16, +6 34 58.82% 91.65% 18 9 / 11 / 22, -2 42 47.62% 33.18% 19 17 / 6 / 17, +11 40 63.75% 98.87% 20 10 / 8 / 18, +2 36 52.78% 67.62% 21 11 / 6 / 17, +5 34 57.35% 88.11% 22 2 / 3 / 5, -1 10 45.00% 34.38% 23 17 / 7 / 12, +10 36 63.89% 97.84% 24 9 / 9 / 22, +0 40 50.00% 50.00% 25 14 / 8 / 28, +6 50 56.00% 89.50% 26 8 / 5 / 13, +3 26 55.77% 78.80% 27 9 / 4 / 9, +5 22 61.36% 91.02% 28 12 / 9 / 21, +3 42 53.57% 73.83% 29 14 / 7 / 15, +7 36 59.72% 93.31% 30 14 / 4 / 20, +10 38 63.16% 99.04% 31 7 / 7 / 24, +0 38 50.00% 50.00% 32 19 / 6 / 19, +13 44 64.77% 99.53% 33 6 / 6 / 8, +0 20 50.00% 50.00% 34 13 / 11 / 20, +2 44 52.27% 65.50% 35 4 / 6 / 10, -2 20 45.00% 27.44% 36 3 / 1 / 6, +2 10 60.00% 81.25% 37 9 / 5 / 22, +4 36 55.56% 84.91% 38 7 / 8 / 11, -1 26 48.08% 40.18% 39 7 / 4 / 11, +3 22 56.82% 80.62% 40 10 / 5 / 15, +5 30 58.33% 89.49% 41 5 / 4 / 11, +1 20 52.50% 62.30% 42 5 / 7 / 6, -2 18 44.44% 29.05% 43 11 / 8 / 17, +3 36 54.17% 74.83% 44 11 / 12 / 17, -1 40 48.75% 41.94% 45 15 / 11 / 22, +4 48 54.17% 77.90% 46 4 / 4 / 12, +0 20 50.00% 50.00% 47 9 / 9 / 20, +0 38 50.00% 50.00% 48 2 / 7 / 11, -5 20 37.50% 5.47% 49 6 / 7 / 17, -1 30 48.33% 39.53% 50 4 / 9 / 19, -5 32 42.19% 8.98% 51 14 / 6 / 20, +8 40 60.00% 96.08% 52 16 / 9 / 19, +7 44 57.95% 91.57% 53 10 / 10 / 26, +0 46 50.00% 50.00% 54 11 / 8 / 19, +3 38 53.95% 74.83% 55 3 / 5 / 4, -2 12 41.67% 25.39% 56 13 / 10 / 19, +3 42 53.57% 72.94% 57 13 / 10 / 21, +3 44 53.41% 72.94% 58 11 / 14 / 21, -3 46 46.74% 27.86% 59 11 / 4 / 23, +7 38 59.21% 96.16% 60 9 / 7 / 12, +2 28 53.57% 68.55% 61 9 / 6 / 25, +3 40 53.75% 77.28% 62 6 / 3 / 17, +3 26 55.77% 82.81% 63 7 / 4 / 13, +3 24 56.25% 80.62% 64 18 / 4 / 14, +14 36 69.44% 99.87% 65 8 / 5 / 9, +3 22 56.82% 78.80% 66 6 / 7 / 15, -1 28 48.21% 39.53% 67 8 / 8 / 18, +0 34 50.00% 50.00% 68 4 / 2 / 10, +2 16 56.25% 77.34% 69 6 / 5 / 11, +1 22 52.27% 61.28% 70 12 / 5 / 13, +7 30 61.67% 95.19% 71 8 / 5 / 9, +3 22 56.82% 78.80% 72 5 / 1 / 4, +4 10 70.00% 93.75% 73 7 / 2 / 13, +5 22 61.36% 94.53% 74 2 / 2 / 6, +0 10 50.00% 50.00% 75 6 / 5 / 11, +1 22 52.27% 61.28% 76 7 / 4 / 9, +3 20 57.50% 80.62% 77 14 / 6 / 22, +8 42 59.52% 96.08% 78 6 / 6 / 12, +0 24 50.00% 50.00% 79 5 / 4 / 7, +1 16 53.13% 62.30% 80 11 / 8 / 15, +3 34 54.41% 74.83% 81 9 / 5 / 16, +4 30 56.67% 84.91% 82 4 / 4 / 8, +0 16 50.00% 50.00% 83 2 / 5 / 19, -3 26 44.23% 14.45% 84 3 / 7 / 18, -4 28 42.86% 11.33% 85 1 / 3 / 12, -2 16 43.75% 18.75% 86 5 / 8 / 13, -3 26 44.23% 21.20% 87 5 / 5 / 10, +0 20 50.00% 50.00% 88 10 / 6 / 14, +4 30 56.67% 83.38% 89 8 / 7 / 17, +1 32 51.56% 59.82% 90 6 / 8 / 6, -2 20 45.00% 30.36% 91 10 / 7 / 15, +3 32 54.69% 75.97% 92 7 / 7 / 24, +0 38 50.00% 50.00% 93 7 / 3 / 12, +4 22 59.09% 88.67% 94 5 / 2 / 7, +3 14 60.71% 85.55% 95 1 / 0 / 5, +1 6 58.33% 75.00% 96 4 / 4 / 6, +0 14 50.00% 50.00% 97 4 / 1 / 13, +3 18 58.33% 89.06% 98 6 / 2 / 16, +4 24 58.33% 91.02% 99 6 / 3 / 5, +3 14 60.71% 82.81% 100 7 / 8 / 11, -1 26 48.08% 40.18% 101 8 / 7 / 17, +1 32 51.56% 59.82% 102 6 / 2 / 10, +4 18 61.11% 91.02% 103 2 / 3 / 7, -1 12 45.83% 34.38% 104 7 / 5 / 16, +2 28 53.57% 70.95% 105 5 / 8 / 7, -3 20 42.50% 21.20% 106 5 / 5 / 10, +0 20 50.00% 50.00% 107 5 / 5 / 12, +0 22 50.00% 50.00% 108 8 / 3 / 9, +5 20 62.50% 92.70% 109 4 / 3 / 5, +1 12 54.17% 63.67% 110 6 / 4 / 18, +2 28 53.57% 72.56% 111 3 / 1 / 4, +2 8 62.50% 81.25% 112 3 / 9 / 8, -6 20 35.00% 4.61% 113 7 / 4 / 15, +3 26 55.77% 80.62% 114 1 / 2 / 1, -1 4 37.50% 31.25% 115 8 / 5 / 9, +3 22 56.82% 78.80% 116 6 / 5 / 9, +1 20 52.50% 61.28% 117 1 / 0 / 1, +1 2 75.00% 75.00% 118 8 / 6 / 20, +2 34 52.94% 69.64% 119 9 / 5 / 8, +4 22 59.09% 84.91% 120 7 / 5 / 20, +2 32 53.13% 70.95% 121 3 / 9 / 4, -6 16 31.25% 4.61% 122 4 / 3 / 3, +1 10 55.00% 63.67% 123 2 / 8 / 10, -6 20 35.00% 3.27% 124 8 / 9 / 15, -1 32 48.44% 40.73% 125 8 / 10 / 16, -2 34 47.06% 32.38% 126 6 / 4 / 16, +2 26 53.85% 72.56% 127 4 / 4 / 6, +0 14 50.00% 50.00% 128 11 / 6 / 13, +5 30 58.33% 88.11% 129 6 / 4 / 18, +2 28 53.57% 72.56% 130 8 / 4 / 22, +4 34 55.88% 86.66% 131 12 / 16 / 38, -4 66 46.97% 22.91% 132 4 / 5 / 11, -1 20 47.50% 37.70% 133 5 / 6 / 7, -1 18 47.22% 38.72% 134 8 / 8 / 22, +0 38 50.00% 50.00% 135 3 / 3 / 8, +0 14 50.00% 50.00% 136 5 / 6 / 9, -1 20 47.50% 38.72% 137 9 / 2 / 15, +7 26 63.46% 98.07% 138 5 / 4 / 7, +1 16 53.13% 62.30% 139 6 / 3 / 9, +3 18 58.33% 82.81% 140 10 / 7 / 21, +3 38 53.95% 75.97% 141 8 / 3 / 11, +5 22 61.36% 92.70% 142 11 / 9 / 18, +2 38 52.63% 66.82% 143 5 / 4 / 9, +1 18 52.78% 62.30% 144 3 / 1 / 6, +2 10 60.00% 81.25% 145 6 / 3 / 5, +3 14 60.71% 82.81% 146 7 / 1 / 8, +6 16 68.75% 98.05% 147 7 / 7 / 14, +0 28 50.00% 50.00% 148 3 / 5 / 10, -2 18 44.44% 25.39% 149 5 / 9 / 10, -4 24 41.67% 15.09% 150 6 / 5 / 25, +1 36 51.39% 61.28% 151 9 / 4 / 5, +5 18 63.89% 91.02% 152 5 / 0 / 7, +5 12 70.83% 98.44% 153 3 / 3 / 4, +0 10 50.00% 50.00% 154 6 / 7 / 13, -1 26 48.08% 39.53% 155 11 / 4 / 9, +7 24 64.58% 96.16% 156 5 / 3 / 12, +2 20 55.00% 74.61% 157 6 / 7 / 9, -1 22 47.73% 39.53% 158 2 / 3 / 5, -1 10 45.00% 34.38% 159 11 / 9 / 14, +2 34 52.94% 66.82% 160 8 / 4 / 12, +4 24 58.33% 86.66% 161 7 / 9 / 10, -2 26 46.15% 31.45% 162 6 / 14 / 16, -8 36 38.89% 3.92% 163 3 / 4 / 7, -1 14 46.43% 36.33% 164 5 / 7 / 10, -2 22 45.45% 29.05% 165 8 / 11 / 13, -3 32 45.31% 25.17% 166 11 / 8 / 17, +3 36 54.17% 74.83% 167 11 / 12 / 29, -1 52 49.04% 41.94% 168 3 / 4 / 9, -1 16 46.88% 36.33% 169 6 / 4 / 14, +2 24 54.17% 72.56% 170 4 / 7 / 9, -3 20 42.50% 19.38% 171 5 / 3 / 12, +2 20 55.00% 74.61% 172 11 / 12 / 21, -1 44 48.86% 41.94% 173 8 / 2 / 12, +6 22 63.64% 96.73% 174 11 / 5 / 14, +6 30 60.00% 92.83% 175 9 / 10 / 17, -1 36 48.61% 41.19% 176 11 / 5 / 14, +6 30 60.00% 92.83% 177 7 / 8 / 21, -1 36 48.61% 40.18% 178 1 / 6 / 11, -5 18 36.11% 3.52% 179 8 / 5 / 15, +3 28 55.36% 78.80% 180 6 / 4 / 2, +2 12 58.33% 72.56% 181 11 / 7 / 10, +4 28 57.14% 82.04% 182 4 / 6 / 16, -2 26 46.15% 27.44% 183 9 / 4 / 21, +5 34 57.35% 91.02% 184 4 / 1 / 9, +3 14 60.71% 89.06% 185 5 / 13 / 12, -8 30 36.67% 3.18% 186 13 / 3 / 12, +10 28 67.86% 99.36% 187 13 / 9 / 24, +4 46 54.35% 79.76% 188 8 / 4 / 10, +4 22 59.09% 86.66% 189 10 / 7 / 11, +3 28 55.36% 75.97% 190 8 / 3 / 7, +5 18 63.89% 92.70% 191 11 / 2 / 23, +9 36 62.50% 99.35% 192 7 / 6 / 11, +1 24 52.08% 60.47% 193 3 / 4 / 9, -1 16 46.88% 36.33% 194 6 / 2 / 8, +4 16 62.50% 91.02% 195 4 / 4 / 14, +0 22 50.00% 50.00% 196 7 / 4 / 7, +3 18 58.33% 80.62% 197 5 / 8 / 15, -3 28 44.64% 21.20% 198 4 / 2 / 6, +2 12 58.33% 77.34% 199 14 / 9 / 15, +5 38 56.58% 84.63% 200 3 / 3 / 8, +0 14 50.00% 50.00% Top Parameters: By LOS [1] par1 64, score 69.44%, LOS 99.870%, Games 36, NetWins +14 [2] par1 32, score 64.77%, LOS 99.532%, Games 44, NetWins +13 [3] par1 186, score 67.86%, LOS 99.364%, Games 28, NetWins +10 [4] par1 191, score 62.50%, LOS 99.353%, Games 36, NetWins +9 [5] par1 30, score 63.16%, LOS 99.039%, Games 38, NetWins +10
Would you mind sharing your CLOP processing program?
regards,
--tom
-
- Posts: 4840
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: learning question about programs
Unfortunately this is only capable of up to 3 parameter types.
Download:
http://www.mediafire.com/download/ej32b ... r-v3.1.rar
Code: Select all
Number of parameters: 3
First parameter: PawnStructure
Second parameter: Mobility
Third parameter: WeakHoles
Param1: Min 30, Max 50
Param2: Min 20, Max 40
Param3: Min 15, Max 35
Total games: 8554
Param1 param2 param3 W / L / D NetW Games Score LOS
30 20 20 0 / 2 / 0, -2 2 0.00% 12.50%
30 20 21 1 / 1 / 0, +0 2 50.00% 50.00%
30 20 31 1 / 0 / 1, +1 2 75.00% 75.00%
30 20 33 1 / 1 / 0, +0 2 50.00% 50.00%
30 20 34 2 / 2 / 0, +0 4 50.00% 50.00%
30 20 35 2 / 1 / 1, +1 4 62.50% 68.75%
30 21 30 0 / 2 / 0, -2 2 0.00% 12.50%
30 21 32 5 / 2 / 1, +3 8 68.75% 85.55%
30 21 34 1 / 2 / 1, -1 4 37.50% 31.25%
30 22 25 1 / 1 / 0, +0 2 50.00% 50.00%
30 22 30 1 / 0 / 1, +1 2 75.00% 75.00%
30 22 32 4 / 1 / 1, +3 6 75.00% 89.06%
30 22 33 1 / 4 / 1, -3 6 25.00% 10.94%
30 22 35 0 / 2 / 0, -2 2 0.00% 12.50%
[...]
50 35 35 1 / 0 / 1, +1 2 75.00% 75.00%
50 36 16 0 / 2 / 2, -2 4 25.00% 12.50%
50 37 15 2 / 0 / 0, +2 2 100.00% 87.50%
50 37 16 2 / 1 / 1, +1 4 62.50% 68.75%
50 37 27 0 / 1 / 1, -1 2 25.00% 25.00%
50 37 28 1 / 0 / 1, +1 2 75.00% 75.00%
50 39 15 0 / 2 / 0, -2 2 0.00% 12.50%
Top Parameters: By LOS
[1] par1 47, par2 22, par3 26 score 100.00%, LOS 99.219%, Games 6, NetWins +6
[2] par1 30, par2 37, par3 19 score 91.67%, LOS 98.438%, Games 6, NetWins +5
[3] par1 44, par2 30, par3 20 score 91.67%, LOS 98.438%, Games 6, NetWins +5
[4] par1 37, par2 24, par3 29 score 87.50%, LOS 98.047%, Games 8, NetWins +6
[5] par1 30, par2 30, par3 25 score 83.33%, LOS 96.875%, Games 6, NetWins +4
http://www.mediafire.com/download/ej32b ... r-v3.1.rar
-
- Posts: 838
- Joined: Thu Jul 05, 2007 5:03 pm
- Location: British Columbia, Canada
Re: learning question about programs
There was some research around 1998 on this (learning evaluation weights) using the program KnightCap and a "TD-Lambda" algorithm.
https://www.google.ca/#q=knightcap+td+lambda
I don't know if that approach would be successful with today's engines.
https://www.google.ca/#q=knightcap+td+lambda
I don't know if that approach would be successful with today's engines.
-
- Posts: 7221
- Joined: Mon May 27, 2013 10:31 am
Re: learning question about programs
At least I know that that approach cost you a lot of eeeleeeectriccityyy.wgarvin wrote:There was some research around 1998 on this (learning evaluation weights) using the program KnightCap and a "TD-Lambda" algorithm.
https://www.google.ca/#q=knightcap+td+lambda
I don't know if that approach would be successful with today's engines.
-
- Posts: 303
- Joined: Sat Apr 28, 2012 6:18 pm
- Location: Austin, TX
Re: learning question about programs
Thank you. A couple of minor questions. Does it handle floating point parameters? And does it handle negative parameters? Regardless of the answers, I'm grateful for your quick response and generosity.Ferdy wrote:Unfortunately this is only capable of up to 3 parameter types.Download:Code: Select all
Number of parameters: 3 First parameter: PawnStructure Second parameter: Mobility Third parameter: WeakHoles Param1: Min 30, Max 50 Param2: Min 20, Max 40 Param3: Min 15, Max 35 Total games: 8554 Param1 param2 param3 W / L / D NetW Games Score LOS 30 20 20 0 / 2 / 0, -2 2 0.00% 12.50% 30 20 21 1 / 1 / 0, +0 2 50.00% 50.00% 30 20 31 1 / 0 / 1, +1 2 75.00% 75.00% 30 20 33 1 / 1 / 0, +0 2 50.00% 50.00% 30 20 34 2 / 2 / 0, +0 4 50.00% 50.00% 30 20 35 2 / 1 / 1, +1 4 62.50% 68.75% 30 21 30 0 / 2 / 0, -2 2 0.00% 12.50% 30 21 32 5 / 2 / 1, +3 8 68.75% 85.55% 30 21 34 1 / 2 / 1, -1 4 37.50% 31.25% 30 22 25 1 / 1 / 0, +0 2 50.00% 50.00% 30 22 30 1 / 0 / 1, +1 2 75.00% 75.00% 30 22 32 4 / 1 / 1, +3 6 75.00% 89.06% 30 22 33 1 / 4 / 1, -3 6 25.00% 10.94% 30 22 35 0 / 2 / 0, -2 2 0.00% 12.50% [...] 50 35 35 1 / 0 / 1, +1 2 75.00% 75.00% 50 36 16 0 / 2 / 2, -2 4 25.00% 12.50% 50 37 15 2 / 0 / 0, +2 2 100.00% 87.50% 50 37 16 2 / 1 / 1, +1 4 62.50% 68.75% 50 37 27 0 / 1 / 1, -1 2 25.00% 25.00% 50 37 28 1 / 0 / 1, +1 2 75.00% 75.00% 50 39 15 0 / 2 / 0, -2 2 0.00% 12.50% Top Parameters: By LOS [1] par1 47, par2 22, par3 26 score 100.00%, LOS 99.219%, Games 6, NetWins +6 [2] par1 30, par2 37, par3 19 score 91.67%, LOS 98.438%, Games 6, NetWins +5 [3] par1 44, par2 30, par3 20 score 91.67%, LOS 98.438%, Games 6, NetWins +5 [4] par1 37, par2 24, par3 29 score 87.50%, LOS 98.047%, Games 8, NetWins +6 [5] par1 30, par2 30, par3 25 score 83.33%, LOS 96.875%, Games 6, NetWins +4
http://www.mediafire.com/download/ej32b ... r-v3.1.rar
regards,
--tom