GeneticAlgorithmA


Requirements for this test


History 2007


Please set rand = Random(123456) before proceeding with GeneticAlgorithmA performance tests.
I think testing with -Xint (no hotspot - interpreter only) gives us more stable results and let us better compare performance.
Done on pcMy

Test No.
 Date/Revision
 Used Time in sec (Time in Properties)
 Used Problem
set
 Startevaluation
 Generation
(on first run/ 2nd+more)
 Evaluation
 JVM Options
 Database
 ConfigurationNo
 Changed Code?
10
 09.10/330
 1. 60
2. 178
 gstpl
 829
 1. 85/n.p.
2. 270
 1. -
2. -
 -Xint
 db4o
 3
 AbstractProperties
9
 07.10/328
 1. 60
2. 60
3. 178
 gstpl
 829
 1. 750/n.possible
2. 86
3. 280
 1. 178
2. -
3. -
 1.-server
2.-Xint
3.-Xint
 db4o
 3
 use of YearSet.isAssignedToAll
8
 06.10/327
 60
 gstpl
 829
 700/not possible
 227
 -server
 db4o
 3
 enhanced YearSet implementations
7
 01.10/318
 63 (60)
 gstpl
 829
 410
 261
 -server
 db4o
 3
 problem: we
changed getBitSet to toBitSet in YearSetImpl1 (rev259 to rev261)
6
 15.08/227
 1. 63 (60)
2. 59
3. 178
 gstpl
 829
 1. 514/650
2. 72
3. 235
 257/230
 1.-server
2.-Xint
3.-Xint
 db4o
 3
 1. embedded server
2. print the evaluation not every generation, but every second.
3. Age of best is 180/200
5
 28.07./178
 60*10 (600)
 gstpl
 825
 4450
 150
 -server
 db4o
 3
 -
4
 28.07./178
 60 (60)
 gstpl
 825 (don't sort db objects so this not the same as in 3)
 495
 239
 -server
 db4o
 3
 -
3
 28.07./178
 60 (60)
 gstpl
 829
 916
 359
 -server
 db4o
 1
 pGaps = 5 (default)
2
 26.07./177
 58 (60)
 lse-f-91
 69522
 55
 53000
 -server
 db4o
 2
 pGaps = 0
1
 15.07./158
 60 (60)
 gstpl
 ca. 830
 980
 309 (WARNING: formula changes, see changed code!!)
 -server
 db4o
 1
 new formula, but for person and room similar changed '2 *':
(1 * (pGaps / allPersons.size() +
                    2 * pMaxGaps)
 5 * (pCollision / allPersons.size() +
                    2 * pMaxCollision))

Configuration No.
 #parents
 #children
 #changed TI's
 mutationRate
 roomChangeRate
 startTimeChangeRate
 dayChangeRate
 bestIndivCorrected?
3 (result from 23+27.07.07 parameter combination tests)
 80
 150
 80
 1
 0.5
 0.6
 80
 yes
2
 8
 50
 10
 10
 10
 10
 100
 yes
1
 8
 50
 10
 10
 1
 10
 5
 yes

30.07

I run the same again (on the same pc75) all results stay but the best and is different:

Number
 Mean Evaluation
 RmsError of Evaluation
 Mean Generation
 #TIs
 #children
 #parents
 mutationRate
 startTime
 roomRate
 dayRate
928
 190
 4
 628
 120
 160
 80
 0.8
 1
 0.1
 90

28.07

With other parameter combination I tried it on pc75 again:
paramCombi-26-07-07.tar.gz

The best result was:

Number
 Mean Evaluation
 RmsError of Evaluation
 Mean Generation
 #TIs
 #children
 #parents
 mutationRate
 startTime
 roomRate
 dayRate
987
 194
 7.1
 665
 80
 160
 80
 0.8
 0.1
 1
 90

Statistics was made over the best 20 results [mean, RmsError]:
Parameter
 Possible Values
 Occured Values
 Mean
 RmsError
 Comment
#children
 80, 100, 120, 140, 160, 180
 100, 120, 160, 180
 154
 26
 140 comes not until 64
80 comes not until 188
mutationRate
 0.1
0.8
1.5
 0.8
1.5
 1.0
 0.3
 0.1 comes not until 41
dayRate
 10, 50, 90
 50, 90
 84
 14
 10 comes not until 150
startTime, (0.1 and 1) #parents (80 and 120), #TI's (80 and 120) have no tendency
For roomRate (0.1, 1, and 10) it seems to be that smaller values are better.

23.07

I run some parameter combinations (See IAlgorithmTest) and here are the results:
paramCombi-23-07-07.tar.gz.

See the column header for more information (The first column 'Number' has no meaning and Mean = Mean(Evaluation)).
The qualitative results are nearly the same if we add an offset (approx. 40) to the slower machines results. You can easily import these files and view them with my nlo project, which uses the same genvlin plotter as gstpl but comes with pnuts as its script engine.

Here are the important parameters:
The parameters: #TI's, startTimeRate, mutationRate and #parents have no tendency.

This is the best result (pc 75) was:
Number
 Mean Evaluation
 RmsError of Evaluation
 Mean Generation
 #TIs
 #children
 #parents
 mutationRate
 startTime
 roomRate
 dayRate
2109
 256
 8.8
 1171
 100
 100
 100
 0.5
 0.5
 0.5
 50.0

The mean values of all good values (evaluation < 340) of every column were.
Number
 Mean Evaluation
 RmsError of Evaluation
 Mean Generation
 #TIs
 #children
 #parents
 mutationRate
 startTime
 roomRate
 dayRate
-
 290 (+-38)
 33 (+-33)
 1336 (+-611)
 39 (+-45)
 84 (+-35)
 37 (+-40)
 17 (+-22)
 16 (+-21)
 10 (+-18)
 31 (+-22)
where the value in brackets was the rms error

If I would include all values not only the good ones we will get the Mean (and RMSError) for


19.07 revision 167

Today I compared sun's jvm with ibms on pcMy. With and without the -server option:
ibmVsSun.jpg



The y axis shows the generation reached in 2 minutes, so higher is better.

15.07 revision 158

Within 1 minute algorithm reaches 189 from 490 with GeneticAlgoProperties.xml == Config,No.1. After generation 800 on pcMy.

13.07. revision 156

no call to _checkGaps() in WeekTISetImpl1._getGaps()
changed new ArrayList<TimeInterval>(getTIs) to getTis()

07.07. revision 153

no explicit countgap and eval started from 490;

21.05.

eval started from 490
4 min 20s: generation 4000 has eval=???
5 min: generation 4480 has eval=172;

06.05.

eval started from 490
5 minutes: generation 4000 has eval=188;
???: generation 4480 has eval=172

14.04.

eval decreases from 976 to 889 in 58 sec (347 generations)
(there was a bug in eval, so beginning eval is far away from 490)