This program combines reinforcement learning (RL) with answer set programming (ASP) to evaluate policies for solving blocks world problems.
The technique used for reinforcement learning is a first-visit Exploring Starts Monte Carlo policy evaluation method.
For the ASP parts, like executing actions and determining the current state of the environment, clingo by Potassco, the Potsdam Answer Set Solving Collection is used.
ASP is also used to generate plans for accelerating the learning process.
Run main.py with python3. Note that this program requires the clingo and numpy modules.
You may want to use a conda environment.
A guide on how to install the clingo Python API can be found here.
Hyperparameters can be set in main.py, specifically at the mc variable.
You can add as many blocks and define the goal state as you wish in *.lp, but performance might decrease as more blocks are used.
To read the full reference and theoretical backgrounds, take a look at and download the thesis.