Robust-optimal-maintenance-planning-through-reinforcement-learning-and-rllib

Code of the paper "POMDP inference and robust solution via deep reinforcement learning: An application to railway optimal maintenance", currently under review.

Requirements

conda create -n rllib python=3.8.13
conda activate rllib
pip install -r requirements.txt

Running

The environment is implemented in env.py. You can modify the hyperparameters of the model and the PPO algorithm in config_*.json and run the training with the command:

export OMP_NUM_THREADS=50; python main.py --model $MODEL

with $MODEL in [belief, gtrxl, lstm]. Alternatively, you can submit the job via bsub command:

bsub < job.bsub

The training will save the average rewards at every evaluation iteration and the best model. You can then run a longer evaluation of the best model by submitting eval.bsub or running evaluation.py.