Speaker
Description
Abstract – Reinforcement Learning (RL) has shown strong potential in optimizing production and manufacturing systems, where complex decisionmaking processes can be modeled as sequential control problems. However, training RL agents in such environments is often computationally expensive and sensitive to stochasticity, leading to unstable learning dynamics and long convergence times. This paper investigates the application of parallelized RL training in a production environment, focusing on its impact on training stability and computational efficiency. By employing vectorized environments and running multiple simulations concurrently, the agent is exposed to a more diverse set of state transitions, reducing variance in policy updates and improving robustness. At the same time, parallelization enables significant reductions in wall-clock training time, providing practical speedup without altering the underlying learning algorithm or the environment. Experimental results on a simulated production system demonstrate that increasing the number of parallel environments leads to more stable learning curves and faster convergence, although diminishing returns appear beyond a certain level of parallelism. The findings highlight the importance of parallel training in scalable RL deployments and provide guidelines for balancing computational resources with learning performance.