How to:
Compile the main program to learn Error Functions
$> cd learn_cp
$> ./compilation/compile_learn_cp.sh
Run the main program to learn Error Functions
/* in learn_cp folder */
$> ./bin/learn_cf_cp (and many options, type -h first to have them)
For instance
/* to learn a error function for AllDifferent-4-5 over 1000 Latin hypercube samplings:*/
$> ./bin/learn_cf_cp -c ad -n 4 -d 5 -l -s 1000
Run the main program to learn Error Functions using precomputed **incomplete** constraint spaces
/* in learn folder */
$> ./bin/learn_cf_cp -i ../learn_ga/spaces/the_file_you_want (plus many other options, type -h first to have them)
For instance
/* to learn a error function from the Ordered-6-9 incomplete constraint space with 100 Latin hypercube samplings:*/
/* 'lt' stands for LessThan, the Ordered constraint implementation we choose for this paper.*/
$> ./bin/learn_cf_cp -c lt -n 6 -d 9 -i ../learn_ga/spaces/lt-6-9-100-la.txt
Run the main program to learn Error Functions using precomputed **complete** constraint spaces
/* in learn folder */
$> ./bin/learn_cf_cp -ci ../learn_ga/spaces/complete_the_file_you_want (plus many other options, type -h first to have them)
For instance
/* to learn a error function from the LinearSum-3-8-12 complete constraint space:*/
/* 'le' stands for LinearEquation.*/
$> ./bin/learn_cf_cp -c le -n 3 -d 8 -p 12 -ci ../learn_ga/spaces/complete_le-3-8-12.txt
Re-run Experiment 1
/* in learn folder */
$> ./run_xp_complete.sh (take about two minutes over 16 cores)
Warning
This will add new results into result files in the results folder. You need to either change the script or rename the corresponding files.
Re-run Experiment 2
/* in learn folder */
$> ./run_xp_incomplete.sh (take about one night over 16 cores)
Warning
This will add new results into result files in the results folder. You need to either change the script or rename the corresponding files.
Re-run Experiment 3
/* in problems/xyz folder */
/* First, compile all models for problem xyz */
$> ./compile_all_xyz.sh
After compiling Sudoku, Killer Sudoku and Magic Square, the fastest
way to run all experiments is to run in the problems folder:
$> ./run_xp_1mon.sh (take about 2 hours over 16 cores)
Analyse result files from Experiments 1 and 2 (this will also print the 3 most frequently learned error functions)
/* in learn_cp folder */
/* Compile first the program printing models */
$> ../learn_ga/compilation/compile_print_model.sh
$> ./analyse_runs.rb results/the_file_you_want
For instance
/* to print results about LinearSum-3-8-12 learned upon the complete constraint space:*/
$> ./analyse_runs.rb results/complete-le-3-8-12-200ms.txt
Analyse result files from Experiment 3
/* in problems folder */
$> ./analyse_results.rb xyz/results/2021-10-27/the_file_you_want
For instance
/* to print results about Sudoku 9x9 with learned error function run through the ICN:*/
$> ./analyse_results.rb results/2021-10-27/3_sudoku_fastforward.txt
Compute the test error of a learned error function over its test set
/* in learn_cp folder */
/* Compile first the program computing the test error */
$> ../lear_ga/compilation/compile_get_score.sh
$> ../learn_ga/bin/get_score (and many options, type -h first to have them)
For instance
/* to get the test error of the most frequently learned error function over the LinearSum-3-8-12 complete constraint space:*/
$> ../learn_ga/bin/get_score -c le -n 100 -d 100 -p 5279 -f 10000000000000000001010000000 -hi ../learn_ga/test_sets/true-le-100-5279-10000_h-m.txt
Folders:
root
- learn_cp: current learning via EFOP
- learn_ga: old learning via GA
- constraints: different constraints to study
- tests: handmade test files (not well-defined unit tests)
- utils: get next configuration, random lib, Latin Hypercube Sampling, ...
problems
- magic_square: Magic Square benchmark
- sudoku: Sudoku benchmarks
- killer_sudoku: Killer Sudoku benchmarks
- vertex_cover: Vertex Cover benchmarks. Work on the small instance, but there is a bug somewhere for the large instance.
explored
- cfn: searching for error functions through a Error Function Network (so meta)
- cfn_cppn: searching for error functions modeled with a CPPN through a Error Function Network
- chebpol: interpolation
- csv: to make CSV files and interpolate error functions from them
- eigen: Eigen lib
- eo: searching for error functions through genetic algorithms
- JLL: everything related to the Johnson–Lindenstrauss Lemma