Genetical tuning

[stegemma]

[...]
This is indeed a very interesting method to try, like creating child from good performing individuals,
and slowly remove individuals that showed decreasing performance. I thought about just giving a million or 2, good test suites instead of game playing.

Something like create individual engines by giving
setoption name passer value 120
setoption name mobility_percent 80
...

Then another individual with different or random values.

Feed the test suite, return the static eval or qsearch score,
the engine that gives the minimum eval error will be the first in the ranking.

Keep a record of all those individual, as after the test those
good performers will be put to a serious game tests.

You can use a set of test positions or a set of PGN games but this way you make your software plays like those of the reference set. of course in a library of PGN there are multiple engines but still I want that my engine has a different playing style. Playing only with its own "children" maybe is dangerous, because that way the software can't use a lot of information that other have, but is for this that I verify the good individual with standard tournament. As other said in this thread, I need a lot of CPU time to get some valid result... but this is an hobby, I don't have a client to satisfy in time

[Ferdy]

I got one and playing around this stuff is interesting. Sample population of 8 and target generation of 20. Fitness of individuals are scored by minimizing the error between the epd score and the individual or engine qsearch score.

Code: Select all

Deuterium GA Eval Opti v1.0

Population         : 8
Target generations : 20
Base engine        : Deuterium v2015.1.35.239 64bit POPCNT

Parameters or genes considered to optimize fitness of individuals:
0, PawnValueEg
1, KnightValueEg
2, BishopValueEg
3, RookValueEg
4, QueenValueEg

Chromosome:
['PawnValueEg', 'KnightValueEg', 'BishopValueEg', 'RookValueEg', 'QueenValueEg']

Create individuals (military ranks) and all starts with random values.

Code: Select all

Create and initialize individuals:
name:    Private, par: [103, 302, 322, 549, 1003], error 1000.00000 cp
name:      Major, par: [117, 320, 322, 546, 1000], error 1000.00000 cp
name:    Captain, par: [116, 317, 304, 508, 1046], error 1000.00000 cp
name:    Colonel, par: [104, 298, 306, 508, 1043], error 1000.00000 cp
name:   Sergeant, par: [106, 317, 307, 544, 1003], error 1000.00000 cp
name: Lieutenant, par: [120, 319, 319, 550, 1026], error 1000.00000 cp
name:    General, par: [108, 311, 304, 502, 1034], error 1000.00000 cp
name:   Corporal, par: [124, 307, 325, 525, 1031], error 1000.00000 cp

Then run each individual to a fitness system, each will analyze around 9000 + positions and returns it average error based from the optimal score in the epd's. The individual will attempt to get close the the ce value.

Code: Select all

8/4k3/6p1/7p/3K1P2/6PP/5n2/6N1 w - h6 c9 "1-0"; sm "d4e3"; ce 134;

Code: Select all

::: Run fitness, at generation 1 :::

engine index 0, name Private
['PawnValueEg', 103]
['KnightValueEg', 302]
['BishopValueEg', 322]
['RookValueEg', 549]
['QueenValueEg', 1003]

executing fitness of Private ... at generation 1
engine id 0, time: 1s, numPos: 9792, aveError: 113.34498

engine index 1, name Major
['PawnValueEg', 117]
['KnightValueEg', 320]
['BishopValueEg', 322]
['RookValueEg', 546]
['QueenValueEg', 1000]

executing fitness of Major ... at generation 1
engine id 1, time: 2s, numPos: 9787, aveError: 113.18147

engine index 2, name Captain
['PawnValueEg', 116]
['KnightValueEg', 317]
['BishopValueEg', 304]
['RookValueEg', 508]
['QueenValueEg', 1046]

executing fitness of Captain ... at generation 1
engine id 2, time: 1s, numPos: 9789, aveError: 113.59996

and so on.

Once completed for this generation number, engines are sorted best perfomer first.

Code: Select all

engine sorted by minimum eval_error:
name:   Corporal, par: [124, 307, 325, 525, 1031], error 113.01696 cp
name: Lieutenant, par: [120, 319, 319, 550, 1026], error 113.08155 cp
name:      Major, par: [117, 320, 322, 546, 1000], error 113.18147 cp
name:    Private, par: [103, 302, 322, 549, 1003], error 113.34498 cp
name:    Colonel, par: [104, 298, 306, 508, 1043], error 113.54211 cp
name:    General, par: [108, 311, 304, 502, 1034], error 113.58431 cp
name:    Captain, par: [116, 317, 304, 508, 1046], error 113.59996 cp
name:   Sergeant, par: [106, 317, 307, 544, 1003], error 113.67875 cp

Corporal was the best with only average error of 113.01696 cp.

Then the crossover thing.
The first crossover (Sergeant is not included in the random selection)

Code: Select all

Crossover name  : first random crossover
first engine    : Colonel
second engine   : General
Crossover the param values of General to Colonel
Crossover starts at index 0 or PawnValueEg
crossover rate : 0.6
Crossing over done!! only param values of Colonel was updated

name:    Colonel, par: [108, 311, 304, 508, 1043], error 1000.00000 cp

And the second crossover (Sergeant is not included in the random selection)

Code: Select all

Crossover name  : second random crossover
first engine    : Corporal
second engine   : Private
Crossover the param values of Private to Corporal
Crossover starts at index 1 or KnightValueEg
crossover rate : 0.6
Crossing over done!! only param values of Corporal was updated

name:   Corporal, par: [124, 302, 322, 549, 1031], error 1000.00000 cp

Note A:
1. Individuals involved in the crossovers are not the same.
2. Unfortunately the best performer was involved in the crossover at second crossover, and being the first selected, some of its genes or parameters (with crossover rate of 0.6 or 60%) were altered by the genes from the Private. How would this affect the performance of the total population?, don't be serious this is GA .

After the crossover we will do something to improve the genes of last individual.
First we will generate new random values and replace the old values. Reset the average error to 1000 cp as well.

Code: Select all

Replace param of last ranked engine (Sergeant) with random values:
existing param values
name:   Sergeant, par: [106, 317, 307, 544, 1003], error 113.67875 cp
new param values after replacement
name:   Sergeant, par: [123, 298, 312, 523, 1034], error 1000.00000 cp

Get the top 1 and top 2 individual and help improve the genes of Sergeant.

Code: Select all

Crossing over of top 1 and 2 engine to the updated last ranked engine (Sergeant):
top 1: Corporal
top 2: Lieutenant

First top 1 first

Code: Select all

Crossover name  : top 1 crossover
first engine    : Sergeant
second engine   : Corporal
Crossover the param values of Corporal to Sergeant
Crossover starts at index 1 or KnightValueEg
crossover rate : 0.6
Crossing over done!! only param values of Sergeant was updated

name:   Sergeant, par: [123, 302, 322, 549, 1034], error 1000.00000 cp

Then the top 2.

Code: Select all

Crossover name  : top 2 crossover
first engine    : Sergeant
second engine   : Lieutenant
Crossover the param values of Lieutenant to Sergeant
Crossover starts at index 0 or PawnValueEg
crossover rate : 0.6
Crossing over done!! only param values of Sergeant was updated

name:   Sergeant, par: [120, 319, 319, 549, 1034], error 1000.00000 cp

Here is what the new Sergeant has become.

Code: Select all

new param values of last ranked engine
name:   Sergeant, par: [120, 319, 319, 549, 1034], error 1000.00000 cp

Note B:
1. But wait a minute, the top 1 engine and that is Corporal was no longer the old Corporal that performed best. Its genes were already altered in Note A, no. 2? Nahhh forget about it, we will continue with the next generation .
Here is the list of individuals after those crossovers (no mutation were done here). These individuals will continue for next generations.

Code: Select all

New values after crossover
name:   Corporal, par: [124, 302, 322, 549, 1031], error 1000.00000 cp
name: Lieutenant, par: [120, 319, 319, 550, 1026], error 113.08155 cp
name:      Major, par: [117, 320, 322, 546, 1000], error 113.18147 cp
name:    Private, par: [103, 302, 322, 549, 1003], error 113.34498 cp
name:    Colonel, par: [108, 311, 304, 508, 1043], error 1000.00000 cp
name:    General, par: [108, 311, 304, 502, 1034], error 113.58431 cp
name:    Captain, par: [116, 317, 304, 508, 1046], error 113.59996 cp
name:   Sergeant, par: [120, 319, 319, 549, 1034], error 1000.00000 cp

Note C:
1. Here the average error of Corporal was resetted to 1000 cp, because it was involved in the crossovers. But in the history of generations that we will see later, at generation 1 the Corporal is best with its genes and average error saved.

So what now, repeat the process but maked it at generation 2.

Code: Select all

::: Run fitness, at generation 2 :::

engine index 0, name Corporal
['PawnValueEg', 124]
['KnightValueEg', 302]
['BishopValueEg', 322]
['RookValueEg', 549]
['QueenValueEg', 1031]

executing fitness of Corporal ... at generation 2
engine id 0, time: 2s, numPos: 9789, aveError: 113.09889

[...]

Until generation 20, our target generation.

Code: Select all

::: Run fitness, at generation 20 :::

engine index 0, name General
['PawnValueEg', 120]
['KnightValueEg', 319]
['BishopValueEg', 322]
['RookValueEg', 549]
['QueenValueEg', 1046]

executing fitness of General ... at generation 20
engine id 0, time: 2s, numPos: 9784, aveError: 112.93091

engine index 1, name Major
['PawnValueEg', 120]
['KnightValueEg', 319]
['BishopValueEg', 322]
['RookValueEg', 549]
['QueenValueEg', 1045]

executing fitness of Major ... at generation 20
engine id 1, time: 1s, numPos: 9784, aveError: 112.93152

[...]

That is it, the General takes over.

Code: Select all

Fitness completed to all individuals at generation 20

engine sorted by minimum eval_error:
name:    General, par: [120, 319, 322, 549, 1046], error 112.93091 cp
name:      Major, par: [120, 319, 322, 549, 1045], error 112.93152 cp
name:    Private, par: [123, 319, 322, 549, 1050], error 112.96238 cp
name: Lieutenant, par: [120, 319, 322, 549, 1028], error 112.97394 cp
name:    Captain, par: [120, 319, 322, 549, 1028], error 112.97394 cp
name:    Colonel, par: [123, 319, 322, 546, 1043], error 112.97894 cp
name:   Corporal, par: [123, 319, 322, 549, 1031], error 112.99785 cp
name:   Sergeant, par: [123, 319, 322, 546, 1034], error 112.99908 cp

Here is the history of top performing individuals, amazing GA, it successfully minimizes the average error slowly.

Code: Select all

History of top performers per generation:
generation   1, name   Corporal, par [124, 307, 325, 525, 1031], error 113.01696
generation   2, name   Sergeant, par [120, 319, 319, 549, 1034], error 113.02238
generation   3, name   Sergeant, par [120, 319, 319, 549, 1034], error 113.02238
generation   4, name    Colonel, par [120, 319, 319, 549, 1043], error 113.00859
generation   5, name    Colonel, par [120, 319, 319, 549, 1043], error 113.00859
generation   6, name    Colonel, par [120, 319, 322, 546, 1043], error 112.93663
generation   7, name    Colonel, par [120, 319, 322, 546, 1043], error 112.93663
generation   8, name    Colonel, par [120, 319, 322, 546, 1043], error 112.93663
generation   9, name   Corporal, par [120, 319, 322, 549, 1031], error 112.95769
generation  10, name   Corporal, par [120, 319, 322, 549, 1031], error 112.95769
generation  11, name   Corporal, par [120, 319, 322, 549, 1031], error 112.95769
generation  12, name   Sergeant, par [120, 319, 322, 549, 1034], error 112.95513
generation  13, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  14, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  15, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  16, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  17, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  18, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  19, name    General, par [120, 319, 322, 549, 1046], error 112.93091
generation  20, name    General, par [120, 319, 322, 549, 1046], error 112.93091

And the other info in summary.

Code: Select all

Top 5 performing individuals at given generation:
[13, 'General', '[120, 319, 322, 549, 1046]', 112.93091]
[14, 'General', '[120, 319, 322, 549, 1046]', 112.93091]
[15, 'General', '[120, 319, 322, 549, 1046]', 112.93091]
[16, 'General', '[120, 319, 322, 549, 1046]', 112.93091]
[17, 'General', '[120, 319, 322, 549, 1046]', 112.93091]

Code: Select all

Population             : 8
generation completed   : 20
Fitness positions      : 10000
Test duration          : 5.6 minutes

Code: Select all

Parameter values after GA tuning:
PawnValueEg   = 120
KnightValueEg = 319
BishopValueEg = 322
RookValueEg   = 549
QueenValueEg  = 1046

Code: Select all

Parameter values default:
option name PawnValueMg type spin default 89 min 70 max 110
option name PawnValueEg type spin default 125 min 90 max 150
option name KnightValueMg type spin default 325 min 300 max 350
option name KnightValueEg type spin default 315 min 300 max 350
option name BishopValueMg type spin default 328 min 300 max 350
option name BishopValueEg type spin default 327 min 300 max 350
option name RookValueMg type spin default 504 min 480 max 530
option name RookValueEg type spin default 498 min 480 max 530
option name QueenValueMg type spin default 996 min 950 max 1200
option name QueenValueEg type spin default 999 min 950 max 1200

The fun starts when tinkering the crossover and mutation methods to improve engine performance.

[stegemma]

This is an AMAZING description of my algorithm!!!

The idea to use your kind of fitness function seems to speed up tremendously and, as you see, in only 20 generations and 8 engines you can get valid results. Maybe with my system of tournaments I need a lot of time but it was born because I would like to make it evaluates parameters while thinking, at runtime in true games. That way it doesn't works but the idea remains the same.

[Ferdy]

Tried it at close to 300k pos as training sets and 30 generations with following results.

Code: Select all

Population         : 8
Target generations : 30
Base engine        : Deuterium v2015.1.35.239 64bit POP

Parameters or genes considered to optimize fitness of individuals:
0, PawnValueEg
1, KnightValueEg
2, BishopValueEg
3, RookValueEg
4, QueenValueEg

Chromosome:
['PawnValueEg', 'KnightValueEg', 'BishopValueEg', 'RookValueEg', 'QueenValueEg']

Create and initialize individuals:
name:    Private, par: [107, 325, 313, 486, 1099], aveErr 1000.00000 cp
name:      Major, par: [131, 323, 306, 495, 1033], aveErr 1000.00000 cp
name:    Captain, par: [105, 316, 335, 492, 1114], aveErr 1000.00000 cp
name:    Colonel, par: [136, 321, 320, 528, 963], aveErr 1000.00000 cp
name:   Sergeant, par: [97, 319, 300, 483, 993], aveErr 1000.00000 cp
name: Lieutenant, par: [123, 335, 310, 480, 974], aveErr 1000.00000 cp
name:    General, par: [108, 332, 332, 507, 972], aveErr 1000.00000 cp
name:   Corporal, par: [99, 338, 306, 524, 1038], aveErr 1000.00000 cp

I generate randomly the initial values based from the min and max values for that parameter.

Code: Select all

::: Run fitness, at generation 1 :::

engine index 0, name Private
par PawnValueEg, value 107, min 90, max 150
par KnightValueEg, value 325, min 300, max 350
par BishopValueEg, value 313, min 300, max 350
par RookValueEg, value 486, min 480, max 530
par QueenValueEg, value 1099, min 950, max 1200

Three threads were used taking 3 unique epd files.

Code: Select all

executing fitness of Private ... at generation 1
Thread 2, time: 19.6s, numPos: 97944, aveErr: 113.99047 cp

Thread 3, time: 19.6s, numPos: 98011, aveErr: 114.45435 cp

Thread 1, time: 19.6s, numPos: 97961, aveErr: 114.88791 cp

Average error: 114.44424 cp

engine index 1, name Major
par PawnValueEg, value 131, min 90, max 150
par KnightValueEg, value 323, min 300, max 350
par BishopValueEg, value 306, min 300, max 350
par RookValueEg, value 495, min 480, max 530
par QueenValueEg, value 1033, min 950, max 1200

executing fitness of Major ... at generation 1
Thread 2, time: 19.1s, numPos: 97891, aveErr: 113.82419 cp
Thread 3, time: 19.1s, numPos: 97929, aveErr: 114.14634 cp


Thread 1, time: 19.1s, numPos: 97900, aveErr: 114.59457 cp

Average error: 114.18837 cp

[...]

History of top performers per generation:

Code: Select all

generation   1, name    Colonel, par [136, 321, 320, 528, 963], aveErr 113.79524
generation   2, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   3, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   4, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   5, name Lieutenant, par [105, 316, 335, 508, 1087], aveErr 113.63099
generation   6, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   7, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   8, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation   9, name    Colonel, par [136, 316, 335, 492, 963], aveErr 113.77916
generation  10, name   Sergeant, par [136, 316, 335, 492, 1059], aveErr 113.70142
generation  11, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  12, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  13, name Lieutenant, par [136, 316, 335, 492, 1040], aveErr 113.69146
generation  14, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  15, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  16, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  17, name    Private, par [111, 316, 335, 492, 1045], aveErr 113.69659
generation  18, name   Corporal, par [136, 316, 335, 513, 1150], aveErr 113.53629
generation  19, name   Corporal, par [136, 316, 335, 513, 1150], aveErr 113.53629
generation  20, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  21, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  22, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  23, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  24, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  25, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  26, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  27, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  28, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation  29, name   Corporal, par [133, 316, 335, 513, 1086], aveErr 113.29757
generation  30, name Lieutenant, par [133, 316, 335, 513, 1040], aveErr 113.28823

Top 5 performing individuals at given generation:

Code: Select all

generation 30, name Lieutenant, par [133, 316, 335, 513, 1040], aveErr 113.28823
generation 29, name   Corporal, par [133, 316, 335, 513, 1086], aveErr 113.29757
generation 20, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation 21, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849
generation 22, name      Major, par [137, 316, 335, 513, 1048], aveErr 113.37849

Test duration          : 01h:18m:05s:744ms

Do the self test games, default is using,

Code: Select all

[125, 315, 327, 498, 999]

The tuned engine is using values from Lieutenant, the best performer after 30 generations.

Code: Select all

[133, 316, 335, 513, 1040]

The tuned engine leads by +78 wins after 2467 games so far, in TC 15s + 100ms inc. Not bad at all.

Code: Select all

TestEngine: D2015.1.35.239.tuned
BaseEngine: D2015.1.35.239
Elo0: -1.50, Elo1: 4.50, alpha: 0.05, beta: 0.05
T: 2467, W: 637, L: 559, D: 1271, Score: 51.6%, NetW: +78
Elo: +10, err95: +/-9, LOS: 0.98816
LLR: 1.84, [-2.94, +2.94]

I am researching on mutation, and this would look good only if I can trace the full history of the individuals like, does it able to reduce the aveErr if value of pawn increases or decreases, same with other params. Once this is known mutation can be introduced on certain parameter.

There are a lot to try also on different kinds of crossover, the one I am using is called single point crossover, with crossover point randomly selected.

Ref.
http://www.obitko.com/tutorials/genetic ... tation.php

[elcabesa]

So you are entering 3 new individuals on each generation?

[Ferdy]

So you are entering 3 new individuals on each generation?

Not exactly, the total number of individuals is the same.
It starts at 8 and it will also end at 8.
But the genes of the three of those individuals are altered thru crossover per generation, if this is what you mean.
For 8 individuals, 3 are going to be altered, this would be increased as population increases.

[stegemma]

[...]
I am researching on mutation, and this would look good only if I can trace the full history of the individuals like, does it able to reduce the aveErr if value of pawn increases or decreases, same with other params. Once this is known mutation can be introduced on certain parameter.

There are a lot to try also on different kinds of crossover, the one I am using is called single point crossover, with crossover point randomly selected.
[...]

This is the same type of crossover that I use: a single point crossing. I don't think that other kind of crossing could works. In effect, the idea of crossing the last one with the best two gives you some kind of two point crossing and because before crossing the last individual has been set casually, you don't need to add any kind of mutation. The first attempt I've made was with a complex crossing and mutation algorithm but it was too hard to configure. In this last version, you don't need to worry about % of mutations, because they are "embedded". As I've said, the random values grow from bottom to up and the good ones grow from up to bottom. The best individual is a little protected but without a true elitism, so that anything could happen.

It seems to me that you are exploring my type of GA more than me... and that's why I like to share any new (strange) idea that I have!

[Ferdy]

It seems to me that you are exploring my type of GA more than me... and that's why I like to share any new (strange) idea that I have!

, I put on hold on mutation, and instead I am trying to observe the effects of increasing the population and genes.
I am trying now N = 14, genes are also increased.
total_players_to_crossover = N/2 = 7, say 6
then crosssover 1x2, 3x4 and 5x6, so I got 3 crossovers there, then add
top 1 and 2 for last placer for a total of 4 crossovers.
In this version I revised the crossover, from crossover point then go to the left of the chromosome, but same method, this is still single point crossover. From my previous post I define crossover point then go the right of chromosome.

Code: Select all

Population         : 14
Target generations : 20
Fitness positions  : 45000
Base engine        : Deuterium v2015.1.35.239 64bit POP

Parameters or genes considered to optimize fitness of individuals:

Code: Select all

 0, PawnValueMg
 1, PawnValueEg
 2, KnightValueMg
 3, KnightValueEg
 4, BishopValueMg
 5, BishopValueEg
 6, RookValueMg
 7, RookValueEg
 8, QueenValueMg
 9, QueenValueEg
10, MidGameMobilityPercent
11, EndGameMobilityPercent
12, OffensivePercent
13, DefensivePercent

Chromosome:
['PawnValueMg', 'PawnValueEg', 'KnightValueMg', 'KnightValueEg', 'BishopValueMg', 'BishopValueEg', 'RookValueMg', 'RookValueEg', 'QueenValueMg', 'QueenValueEg', 'MidGameMobilityPercent', 'EndGameMobilityPercent', 'OffensivePercent', 'DefensivePercent']

Create and initialize individuals:

Code: Select all

name:       Manila, par: [105, 93, 337, 340, 322, 349, 505, 519, 1131, 982, 193, 121, 260, 297], aveErrCp 1000.00000
name:    Amsterdam, par: [92, 143, 347, 323, 319, 308, 524, 512, 1161, 1149, 158, 157, 251, 287], aveErrCp 1000.00000
name:       London, par: [78, 110, 326, 343, 312, 307, 503, 528, 1062, 962, 196, 179, 226, 204], aveErrCp 1000.00000
name: WashingtonDC, par: [90, 98, 311, 301, 300, 343, 490, 502, 1102, 1191, 189, 193, 152, 242], aveErrCp 1000.00000
name:         Oslo, par: [93, 90, 329, 339, 346, 314, 521, 486, 984, 1182, 118, 195, 263, 294], aveErrCp 1000.00000
name:       Berlin, par: [79, 136, 323, 321, 325, 303, 524, 502, 1110, 1076, 150, 132, 128, 294], aveErrCp 1000.00000
name:     Brussels, par: [83, 146, 329, 350, 350, 304, 502, 500, 1020, 1062, 137, 108, 277, 275], aveErrCp 1000.00000
name:       Warsaw, par: [84, 100, 301, 321, 327, 310, 521, 510, 955, 1008, 188, 161, 267, 250], aveErrCp 1000.00000
name:       Moscow, par: [72, 112, 314, 317, 314, 333, 480, 480, 1147, 1088, 143, 107, 284, 255], aveErrCp 1000.00000
name:        Tokyo, par: [78, 134, 348, 309, 346, 308, 507, 496, 1016, 1039, 166, 121, 194, 299], aveErrCp 1000.00000
name:         Rome, par: [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 171, 93, 177, 196], aveErrCp 1000.00000
name:      Beijing, par: [92, 118, 350, 322, 309, 308, 528, 514, 1063, 1121, 138, 130, 293, 226], aveErrCp 1000.00000
name:        Paris, par: [106, 94, 330, 310, 307, 309, 514, 501, 974, 1127, 112, 102, 151, 135], aveErrCp 1000.00000
name:    Singapore, par: [73, 129, 350, 342, 340, 333, 519, 503, 1133, 1100, 178, 90, 159, 281], aveErrCp 1000.00000

Sample crossover system after the first generation.

Code: Select all

Fitness completed to all individuals at generation 1

engine sorted by minimum eval aveErrCp:
name:        Paris, par: [106, 94, 330, 310, 307, 309, 514, 501, 974, 1127, 112, 102, 151, 135], aveErrCp 116.94172
name:         Rome, par: [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 171, 93, 177, 196], aveErrCp 117.46022
name: WashingtonDC, par: [90, 98, 311, 301, 300, 343, 490, 502, 1102, 1191, 189, 193, 152, 242], aveErrCp 119.11082
name:       Berlin, par: [79, 136, 323, 321, 325, 303, 524, 502, 1110, 1076, 150, 132, 128, 294], aveErrCp 119.50867
name:    Singapore, par: [73, 129, 350, 342, 340, 333, 519, 503, 1133, 1100, 178, 90, 159, 281], aveErrCp 119.85174
name:        Tokyo, par: [78, 134, 348, 309, 346, 308, 507, 496, 1016, 1039, 166, 121, 194, 299], aveErrCp 120.10575
name:       London, par: [78, 110, 326, 343, 312, 307, 503, 528, 1062, 962, 196, 179, 226, 204], aveErrCp 120.83198
name:       Warsaw, par: [84, 100, 301, 321, 327, 310, 521, 510, 955, 1008, 188, 161, 267, 250], aveErrCp 121.67594
name:       Moscow, par: [72, 112, 314, 317, 314, 333, 480, 480, 1147, 1088, 143, 107, 284, 255], aveErrCp 122.51106
name:       Manila, par: [105, 93, 337, 340, 322, 349, 505, 519, 1131, 982, 193, 121, 260, 297], aveErrCp 122.54671
name:         Oslo, par: [93, 90, 329, 339, 346, 314, 521, 486, 984, 1182, 118, 195, 263, 294], aveErrCp 122.57350
name:     Brussels, par: [83, 146, 329, 350, 350, 304, 502, 500, 1020, 1062, 137, 108, 277, 275], aveErrCp 122.68520
name:      Beijing, par: [92, 118, 350, 322, 309, 308, 528, 514, 1063, 1121, 138, 130, 293, 226], aveErrCp 123.00229
name:    Amsterdam, par: [92, 143, 347, 323, 319, 308, 524, 512, 1161, 1149, 158, 157, 251, 287], aveErrCp 123.55571

Crossover name  : random crossover 1
first engine    : London
second engine   : Manila
Crossover the param values of Manila to London
Crossover point at index 1 or PawnValueEg
crossover rate : 0.1
Crossover done!! only param values of London was updated

name:       London, par: [105, 93, 326, 343, 312, 307, 503, 528, 1062, 962, 196, 179, 226, 204], aveErrCp 1000.00000

Crossover name  : random crossover 2
first engine    : Beijing
second engine   : Singapore
Crossover the param values of Singapore to Beijing
Crossover point at index 12 or OffensivePercent
crossover rate : 0.7
Crossover done!! only param values of Beijing was updated

name:      Beijing, par: [92, 118, 350, 342, 340, 333, 519, 503, 1133, 1100, 178, 90, 159, 226], aveErrCp 1000.00000

Crossover name  : random crossover 3
first engine    : Warsaw
second engine   : Tokyo
Crossover the param values of Tokyo to Warsaw
Crossover point at index 5 or BishopValueEg
crossover rate : 0.4
Crossover done!! only param values of Warsaw was updated

name:       Warsaw, par: [78, 134, 348, 309, 346, 308, 521, 510, 955, 1008, 188, 161, 267, 250], aveErrCp 1000.00000

Replace param of last ranked engine (Amsterdam) with random values:
existing param values
name:    Amsterdam, par: [92, 143, 347, 323, 319, 308, 524, 512, 1161, 1149, 158, 157, 251, 287], aveErrCp 123.55571
new param values after replacement
name:    Amsterdam, par: [70, 108, 337, 304, 306, 312, 507, 481, 996, 1046, 147, 136, 107, 229], aveErrCp 1000.00000

Crossover top 1 and 2 for last ranked (Amsterdam):
top 1: Paris
top 2: Rome
Crossover name  : top 1 crossover
first engine    : Amsterdam
second engine   : Paris
Crossover the param values of Paris to Amsterdam
Crossover point at index 10 or MidGameMobilityPercent
crossover rate : 0.7
Crossover done!! only param values of Amsterdam was updated

name:    Amsterdam, par: [106, 94, 330, 310, 307, 309, 514, 501, 974, 1127, 112, 136, 107, 229], aveErrCp 1000.00000

Crossover name  : top 2 crossover
first engine    : Amsterdam
second engine   : Rome
Crossover the param values of Rome to Amsterdam
Crossover point at index 2 or KnightValueMg
crossover rate : 0.1
Crossover done!! only param values of Amsterdam was updated

name:    Amsterdam, par: [80, 123, 334, 310, 307, 309, 514, 501, 974, 1127, 112, 136, 107, 229], aveErrCp 1000.00000

The top performers at generation 20 or last generation.

Code: Select all

Fitness completed to all individuals at generation 20

engine sorted by minimum eval aveErrCp:
name:       Berlin, par: [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
name:        Paris, par: [80, 123, 320, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 114.98762
name:       Moscow, par: [80, 123, 320, 308, 319, 328, 500, 523, 1194, 1151, 94, 133, 135, 135], aveErrCp 115.09147
name:       Warsaw, par: [80, 123, 320, 308, 319, 328, 500, 523, 1194, 1151, 94, 133, 135, 135], aveErrCp 115.09147
name:      Beijing, par: [80, 123, 320, 308, 319, 328, 497, 529, 1001, 975, 140, 122, 131, 187], aveErrCp 115.79937
name:        Tokyo, par: [80, 123, 320, 308, 319, 328, 497, 529, 1001, 975, 140, 183, 212, 134], aveErrCp 116.35492
name:    Singapore, par: [80, 123, 320, 323, 324, 325, 500, 510, 1057, 1195, 92, 197, 109, 220], aveErrCp 116.46940
name:         Oslo, par: [80, 123, 320, 308, 319, 328, 497, 529, 1001, 975, 140, 102, 151, 203], aveErrCp 116.64816
name:         Rome, par: [80, 123, 320, 308, 319, 328, 500, 523, 1194, 1151, 94, 133, 135, 196], aveErrCp 116.72120
name: WashingtonDC, par: [80, 123, 334, 323, 319, 328, 497, 529, 1001, 975, 140, 90, 159, 242], aveErrCp 118.15606
name:    Amsterdam, par: [80, 123, 320, 308, 324, 325, 485, 529, 1060, 1083, 148, 124, 200, 237], aveErrCp 118.30308
name:       London, par: [80, 123, 334, 323, 324, 348, 525, 526, 1046, 984, 171, 108, 286, 134], aveErrCp 118.31431
name:       Manila, par: [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 155, 259], aveErrCp 118.41741
name:     Brussels, par: [80, 123, 334, 323, 324, 348, 525, 526, 1046, 984, 171, 108, 286, 143], aveErrCp 118.44284

History of top performers per generation:

Code: Select all

generation   1, name        Paris, par [106, 94, 330, 310, 307, 309, 514, 501, 974, 1127, 112, 102, 151, 135], aveErrCp 116.94172
generation   2, name        Paris, par [106, 94, 330, 310, 307, 309, 514, 501, 974, 1127, 112, 102, 151, 135], aveErrCp 116.94172
generation   3, name        Paris, par [93, 90, 329, 339, 346, 314, 521, 486, 984, 1182, 112, 102, 151, 135], aveErrCp 116.61102
generation   4, name        Paris, par [93, 90, 329, 339, 346, 314, 521, 486, 984, 1182, 112, 102, 151, 135], aveErrCp 116.61102
generation   5, name        Paris, par [93, 90, 329, 339, 346, 314, 521, 486, 984, 1182, 112, 102, 151, 135], aveErrCp 116.61102
generation   6, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation   7, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation   8, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation   9, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  10, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  11, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  12, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  13, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  14, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  15, name        Paris, par [80, 123, 334, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 115.14098
generation  16, name       Berlin, par [105, 123, 320, 308, 319, 328, 497, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 115.10785
generation  17, name        Paris, par [80, 123, 320, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 114.98762
generation  18, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
generation  19, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
generation  20, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328

Top 5 performing individuals at given generation:

Code: Select all

generation 18, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
generation 19, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
generation 20, name       Berlin, par [80, 123, 320, 308, 319, 328, 500, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 114.54328
generation 17, name        Paris, par [80, 123, 320, 323, 324, 325, 485, 529, 1060, 1128, 112, 102, 151, 135], aveErrCp 114.98762
generation 16, name       Berlin, par [105, 123, 320, 308, 319, 328, 497, 529, 1001, 975, 140, 102, 151, 122], aveErrCp 115.10785

Test duration          : 00h:17m:17s:862ms

Now look at the history, the average error has been consistently hammered down by Paris and Berlin .

[elcabesa]

So you are entering 3 new individuals on each generation?

Not exactly, the total number of individuals is the same.
It starts at 8 and it will also end at 8.
But the genes of the three of those individuals are altered thru crossover per generation, if this is what you mean.
For 8 individuals, 3 are going to be altered, this would be increased as population increases.

Yes,I mean replace/alter ☺

[elcabesa]

Just 2 questions:
1 have you written you own framework for GA ?
2 how do you calculate fitness? Eval, quiescent search or other?
3 what about elo verification?