A Supervised Learning Approach to Robust Reinforcement Learning for Job Shop Scheduling

A three-step-approach that combines RL and Supervised Learning techniques. The initially trained RL policy is used as a labelling oracle that generates state-action pairs which are then augmented with varying permutation percentages to transpose job orders. These state-action pairs serve as data sets for re-training models in a supervised learning setup that uses Dropout layers to improve robustness.

Usage

• python -m src.main --help

Requirements

• Stable-baselines3: https://github.com/DLR-RM/stable-baselines3
• OpenAI gym
• OpenAI gym box2d
• Swig (requirement to install box2d)
• An OpenAi Gym environment for the Job Shop Scheduling problem: https://github.com/prosysscience/JSSEnv