Challenge — Autonomous Control for Complex Systems in Dynamic Environments
Traditional control systems often operate suboptimally, especially under dynamic conditions. As more sensors and input data are introduced, controller design and maintenance become increasingly complex. In real-world applications, controllers must meet strict stability requirements, and their behavior needs to remain comprehensible for operators and engineers.
Solution — Model-Based Reinforcement Learning with Adaptive Plant Modeling
We use model-based reinforcement learning with a supervised-trained deep neural network that models the controlled plant. The model is updated at regular intervals to learn new control strategies as conditions change. The optimization criteria (reward function) can be adjusted by the operator on the fly—even after training—enabling flexible objectives while maintaining a structured control approach.
Impact — Lower Control Error, Less Manual Control & Scalable Energy Savings
- 60% decrease in the error metric for controller quality
- Autonomous control can be scaled to other plants, enabling significant energy savings
- Manual plant control reduced from 20 days to < 1 day
- Autonomous control is positioned to disrupt how the processing industry approaches advanced process control