The code is tested on Linux with a CPU cluster and it will be a further improvement for running with GPU. There are two ways provided to build the environment for running the code.

• All the packages (listed in "requirements.txt") are installed based on python3.6. After it is done, install ray with "sudo pip install ray". In addition, install Jacinle if needed.
• Run the docker file ("Dockerfile") to setup the environment.

• "NLM_MBRL.py" is the entry of running this code.
• "NLM_MBRL_config.py" is the main file with the hyper-parameter settings.
• "models.py" is the main file for constructing the model.

This repo contains 3 graph tasks and 5 family tree tasks under the multi-task reasoning setting.

• An example of the command for training is "jac-run NLM_MBRL.py".
• "model.weights" in the folder of "saved_model" is the pre-trained model weights. An example of testing is "jac-run NLM_MBRL.py --task_mode=test".

• Neural Logic Machines
• MuZero