Sensitivity analysis of EncodedRL : Assessing the impact of state compression levels on dynamic scheduling performance
David Heik, Fouad Bahrpeyma, Dirk Reichelt · 2025
Modern industrial environments face the challenge of enhancing efficiency while managing increasing complexity and uncertainty.Reinforcement learning (RL) offers a promising approach to adaptive decision-making.A fundamental prerequisite is the use of simulation.This provides a safe, controllable, and reproducible environment where agents can explore, learn from failure, and adapt to dynamic production conditions without risking real-world disruptions.However, deploying RL in production remains challenging due to complex problem spaces, large neural networks, and high computational costs.This study builds on EncodedRL, a method that combines unsupervised state space compression via autoencoders with multiagent RL to reduce model complexity while retaining decision-relevant information.While prior results were encouraging, the impact of compression levels on policy quality remained unclear.The present study addresses this gap through systematic experimentation.Results indicate that moderate compression significantly reduces training time while improving scheduling performance.These findings support scalable, resource-efficient RL applications in industrial settings.