README.md

Monte Carlo Tree Search on Perfect Rectangle Packing Problem Instances

Dataset

The dataset of 1000 Guillotinable and 1000 Braam rectangle packing problems with 20 tiles can be found in problems.

Installation

pip3 install requirements.txt

To run experiments, you will need to have a PostgreSQL database running.

After installing the PostgreSQL database you can create one by logging into the psql shell and running:

postgres=# create role binpack_user;
CREATE ROLE
postgres=# create database binpack;
CREATE DATABASE
postgres=# grant all on database binpack to binpack_user;
GRANT
postgres=# alter role binpack_user with login;
postgres=# alter user binpack_user with password 'password';

Copy the file visapi/secrets.json.local to visapi/secrets.json:

cp visapi/secrets.json.template visaip/secrets.json

Then edit it with a secret key and the details to connect to the database:

The database values meanings are in order:
HOST, DATABASE_NAME, USER, PASSWORD, PORT

For the example provided, the contents of the file would be:

{
    "SECRET_KEY": "something_random",
    "DATABASE": ["localhost", "binpack", "binpack_user", "password", ""]
}

To create a database schema, you will need to run:

python3 manage.py migrate

All experiments results will be saved to the database, in the table affinity.

Access them using the django shell:

python3 manage.py shell_plus

And then:

Result.objects.all()

See Django ORM reference.

Or access them using postgresql shell:

select * from affinity;

All the problem results will be available in the database with all the metrics needed to finish them (number of tiles placed etc.), and their scores (success / failure).

Running experiments

MCTS algorithm

Run:

PYTHONPATH=engine/ python3 manage.py run_mcts

to see all the options for the runner.

Example:

Run dynamically generated problems:

PYTHONPATH=engine/ python3 manage.py run_mcts 20 11 11

From file:

PYTHONPATH=engine/ python3 manage.py run_mcts 20 11 11 --from_file

Neural Network algorithm

This algorithm performs only one tile placement. To get the final solution calls the NN predict multiple times.

To train the neural network you can use:

PYTHONPATH=engine/ python3 manage.py just_train