Neural Logic Reinforcement Learing

Implementaion of Neural Logic Reinforcement learning and several benchmarks. Neural Logic Reinforcement Learning uses deep reinforcement leanring methods to train a differential indutive logic progamming architecture, obtaining explainable and generalizable policies. Paper accepted by ICML2019.

Enviornments

Developed in python2.7, Linux enviornment.

Dependencies

• Create a virtual env
• numpy: downgrade to numpy v 1.14.0 using "python -m pip install numpy=1.14.0" in your virtual env
• tensorflow

User Guide

• use main.py to run the experiments
• --mode= to specify the running mode, can be "train" or "generalize", where generalize means to run a generalization test.
• --task= to specify the task, can be "stack", "unstack", "on" or "cliffwalking".
• --algo to specify agent type, can be "DILP", "NN" or "Random"
• --name to specify the id of this run.
• for example: python main.py --mode=train --algo=DILP --task=unstack --name=ICMLtest