An implementation of the ARNOLD algorithm for learning polynomial constraints from positive data.

• scipy
• numpy
• math
• itertools
• copy
• argparse
• csv

• minizinc
• pymzn
• pickle
• random

Install all the packages mentioned above, unzip the code in any directory with write permissions and you should be able to run the code.

To learn constraint call run.py with the following arguments:

1. --num_sum
2. --num_prod
3. --negation
4. --negZ <should <= inequality be considered>
5. --folder
6. --file_name
7. --example (use multiple times to input multiple examples)
8. --input_var
9. --var

Most of the input arguments have a default value in the code.

Here is one example to call run.py: python run.py --var W F demand production ship --input_var W F demand production --example W=4 F=3 demand=[30,20,35,20] production=[40,40,25] ship=[[15,0,16,9],[0,20,9,11],[15,0,10,0]]

It will create a minizinc file in the specified folder or the location of the code if not specified.

Currently the implementation is in the experiment mode. Given any minizinc file as an input, the algorithm samples some examples and learns a model which is then written in a minizinc file. To run an experiment follow these steps:

1. Make sure there is a folder inside "experiments/" with the name of the problem, for example shipping.
2. Inside shipping folder, there should be 3 files:
   • "shipping.mzn" : this is the model we try to learn
   • "shipping.mzn.vars" : this contains the name of all the variables and constants in the model
   • "shipping.mzn.inputvars" : this contains the name of all the constants in the model (Please have a look at some of the folders which has already been put there to run experiments)
3. Now to run experiments for a particual problem simply call "experiment.py" with the following arguments: --file --sum --prod

One example to run an existing problem in the experiment would be to run the following command:

python experiment.py --file shipping --sum 1 --prod 1 2 3

This command will run the experiments for shippping problem be fixing number of terms to 1 and varying the number of variables in each term. It will generate the results of the experiment with values on recall, precision and time inside "experiments/shipping/results/". The resulted minizinc file for each experiment is stored in "experiments/shipping/minizinc/"

Mohit Kumar < mohit dot kumar at cs dot kuleuven dot be >