Workshop on Practical Insights into RL for Real Systems

RLC 2025

August 5th, 2025

Edmonton, Canada

State-of-the-art reinforcement learning (RL) methods have demonstrated their ability to solve difficult and complex problems in areas such as video games and natural language processing. These successes suggest that RL algorithms are powerful enough to tackle other complex problems, including the control and optimization of real-world systems. Some of these systems, like supply chains and power grids, are currently managed by well established approaches. However, these methods can become suboptimal when faced with issues of scale, uncertainty, and complexity. Other systems, such as nuclear fusions reactors or autonomous robots, still lack sufficiently powerful control algorithms. In both cases, RL presents a promising solution. However, there are still many challenges to applying RL for such systems, including issues related to performance, evaluation, and certifiability, which depend on the nature and context of each problem. In this workshop, we aim to answer the question: Under which conditions can RL be the right solution for a real system? Unlike previous workshops on deploying RL in the real world, our focus of this workshop will not only be on technical approaches to address these issues, although they are welcome. Instead, we aim to provide a space for RL practitioners, who have chosen RL as the path for their problems, to discuss their experiences and offer practical insights into the implementation of RL in such settings. We plan to address five key questions:

When should we use RL? For which problem characteristics is RL more relevant than other approaches? These considerations could include problem assumptions, mathematical considerations, and deployment requirements.
What methodology should be used to evaluate the RL agent's performance and robustness towards real world deployment, and compare it with other approaches?
How to ensure compliance with the application requirements? What are the challenges and trade-offs in terms of performance, complexity, and robustness of existing approaches?
How to develop a common language to communicate effectively with decision makers and experts working on these systems?
What other practical challenges should practitioners consider when aiming to implement RL in real systems?