Language-Free Acq

Learning constraint networks over unknown constraint languages

We propose a method to efficiently solve the Language-Free Acq optimisation problem by giving an optimal consistent constraint network, if it exists. We have implemented our method in the Python programming language. Our program take as input a training set in the form of a file with one assignment by line completed by the classification for this assignment (solution or non-solution). We compute the vocabulary and generate the corresponding Partial Max-Sat instance and solve it using the OR-Tools library.

Requirements

tqdm~=4.66.3
ortools~=9.10.4067

Tqdm is a progress bar library, used to display the progress of the algorithm.

Installation

pip install -r requirements.txt

Usage

In this new version of the program, the solver is OR-Tools. You don't have to specify the solver anymore.

Use the program in your own code

Install the program with pip: pip install languageFreeAcq

You can now use it in your own code:

from src.languageFreeAcq import Acquisition, CspScopesRelations


lfa = Acquisition()
csp: CspScopesRelations = lfa.learn(
  file_train=_examples_file_path_, max_examples=_maximum_number_examples_, timeout=_timeout_)
problem_variables = list(range(0, lfa.get_variables_numbers()))
problem_domain = lfa.get_domains()
relations_learned = csp.get_relations(0)  # The first learned relation
tuples_learned = csp.get_scope(0)  # The scope where the first learned relation is applied to form
# the constraint network

Edit this program

You can use the script main.py with the following commands:

• main.py all to run the program on all the datasets of the paper.
• main.py dataset_name to run the program on a specific dataset.
• main.py custom to run the program on the custom dataset. You can configure this by modifying the experiments/xp_custom.py file.

Reproduce the experiments of the paper

The paper associated with this program is:

Christian Bessiere, Clément Carbonnel, Areski Himeur:
Learning Constraint Networks over Unknown Constraint Languages. IJCAI 2023

Because the program now use a new solver, the results may be slightly different from the ones in the paper. Moreover, there is some difference between the program in this repository and the one used in the paper notably the way indexes are used. In the paper, the indexes are 1-based while in this repository, the indexes are 0-based. This difference is due to the fact that the solver used in the paper was the Max-Sat solver from the Minisat library which uses 1-based indexes.

The original code can be found here: https://gite.lirmm.fr/coconut/language-free-acq

---

To reproduce the experiments of the paper, you can use the script main.py all which will run the program on all the datasets of the paper. You can also run the program on a specific dataset by using the command main.py dataset_name.

The results appear in the console with the following format:

FILE_TRAIN: data/_.csv  NB_EXAMPLES: _  FILE_TEST: data/_.csv  KR: (_, _) ACCURACY: _ RELATION: _  NETWORK: _ TIME: _

where accuracy is the accuracy of the network on the test set, relation is a boolean indicating if the network obtained has the same relation as the one used to generate the dataset, network indicates if the network is strictly the same as the one used to generate the dataset.