richoux / QUBO_learning

Learning QUBO matrices with GHOST

for one-hot encoding only (ICCS 2023 paper).
for unary and one-hot encodings.

Learning QUBO matrix from data

This program aims to learn pattern composition representing a Q matrix from data.

Project tree

• bin: where the project binaries are generated
• include: headers of third-party libraries
• lib: third-party libraries
• obj: where the project object files are generated
• results: contains the results of experiments run for our paper.
• scripts: scripts to run different experiments in the paper.
• spaces: training and test sets. spaces/complete and spaces/incomplete contain training sets of Experiment 1 and 2, respectively.
• src: source code of the project. src/make_spaces contains the source code to generate training and test sets. src/models contains the code of our different models. The model presented in the paper is src/models/block_learning_opt (and src/models/block_learning_sat for the model without the objective function). src/explored contains different models and learning schemes that have been tested before src/models/block_learning_opt and src/models/block_learning_sat.

How to:

Compile programs

$> make

If you need details of the solving process, then type

$> make MYFLAGS=-DGHOST_BENCH

Run the main program to learn a Q matrix

$> ./bin/learn_q_opt (and many options, type -h first to have them)`

For instance

Learn a q matrix for AllDifferent-4-4:

$> ./bin/learn_q_opt -f spaces/complete/alldiff-4_4.txt

Learn a q matrix for AllDifferent-4-4 and write q in an output file:

$> ./bin/learn_q_opt -f spaces/complete/alldiff-4_4.txt -m matrix_file.txt

Learn a q matrix for AllDifferent-4-4 and write the pattern composition vector in an output file:

$> ./bin/learn_q_opt -f spaces/complete/alldiff-4_4.txt -r result_file.txt

The one-hot encoding is applied by default. To choose the unary encoding, use the option -e 1

Run the main program to learn a Q matrix with weak learners

$> ./bin/weak_learn_q_opt (and many options, type -h first to have them)`

For instance

Learn a q matrix for AllDifferent-4-4 with 5 weak learners:

$> ./bin/weak_learn_q_opt -f spaces/complete/alldiff-4_4.txt -w 5

Learn a q matrix for AllDifferent-12-12 with 7 weak learners over the same samples for each weak learners (here, a pre-sampled set of 4 solutions and 4 non-solutions):

$> ./bin/weak_learn_q_opt -f spaces/incomplete/alldiff-12_12_4.txt -w 7

Learn a q matrix for AllDifferent-12-12 with 9 weak learners over 10 random samples (different samples for each weak learners):

$> ./bin/weak_learn_q_opt -f spaces/incomplete/alldiff-12_12_10000.txt -w 9 -n 10

Run programs to build spaces

Generate complete spaces (scan all possible candidates):

$> ./bin/make_complete_spaces (and many options, type -h first to have them)

Generate incomplete spaces (draw positive and negative candidates):

$> ./bin/make_incomplete_spaces (and many options, type -h first to have them)

Generate test spaces (build positive candidates and draw negative ones):

$> ./bin/make_test_spaces (and many options, type -h first to have them)

For instance

Create the file alldiff.txt containing the complete space of AllDifferent over 4 variables with domains of size 6:

$> ./bin/make_complete_spaces -c ad -n 4 -d 6 -o alldiff.txt

Create the file linear_equation.txt containing an incomplete space of the LinearSum x+y+z=10 with domains of size 5, by sampling 1,000 positive and 1,000 negative candidates:

$> ./bin/make_incomplete_spaces -c le -n 3 -d 5 -p 10 -s 1000 -o linear_equation.txt

Create the file channel.txt containing a test space of Channel over 20 variables with domains of size 20, with 50 positive and 50 negative candidates:

$> ./bin/make_test_spaces -c ch -n 20 -d 20 -s 50 -o channel.txt

Warning

Running experiments with scripts in the scripts folder will generate new result files in the results folder.