Reproduce “Heuristic Learning for Active Flow Control Using Coding Agents” on OpenResearch

About this paper

Paul Garnier, Jonathan Viquerat, Elie Hachem

This research investigates whether modern coding agents can autonomously discover explicit feedback controllers for active flow control problems that are competitive with state-of-the-art Deep Reinforcement Learning (DRL) policies. The study demonstrates that these heuristically learned controllers can match or surpass DRL performance in 10 out of 13 benchmarks while offering full inspectability and improved transferability.

Key insights

• Modern coding agents are capable of autonomously discovering non-trivial explicit feedback controllers for active flow control that are competitive with DRL, without requiring adjoint solvers or differentiable simulation chains.
• Explicit, human-readable control laws provide critical interpretability, allowing for direct inspection of feedback structures, gains, and delays, which enhances physical understanding and trust in the control mechanisms.
• Heuristic learning exhibits higher sample efficiency in early optimization stages and facilitates more effective transfer learning through adaptable source code, contrasting with the limitations of opaque DRL policies.

Summary by alphaXiv.