YANN-RL slashes training time for chemical process control

Researchers Austin Braniff and Yuhe Tian have introduced YANN-RL, a reinforcement learning framework that leverages Y-wise Affine Neural Networks to address longstanding trust and training challenges in chemical process control. By strategically initializing actor and critic networks, YANN-RL provides confident, interpretable starting points that drastically cut training time and data requirements. This is a critical advance because traditional RL algorithms like PPO, SAC, DDPG, and TD3 have struggled to gain adoption in industrial settings due to slow convergence and opaque decision-making.

In benchmark tests against nonlinear model predictive control (NMPC) across three publicly available case studies from the PC-Gym library—a continuous stirred tank reactor, a four-tank system, and a multistage extraction column—YANN-RL approached NMPC’s performance without needing a full nonlinear model. This means operators can deploy RL control faster and with greater confidence. The work, accepted at the 23rd IFAC World Congress 2026, marks a practical step toward widespread RL adoption in chemical and process industries.