To date industry relies on classical controllers (PID, feedforward, rule-based), which are tuned manually by engineers relying on insight and experience. Automatic tuning can be used for simple systems, but when complexity of the dynamics and variability of conditions increases manual tuning by experts is still the only way. Alternatively, model-based control can be used, but for these complex systems modelling is limited in accuracy due to the effort and knowledge required resulting in insufficient performance on real systems.

To address these limitations, data-driven AI methods – like Reinforcement Learning (RL) – are being developed that learn directly from machine data. However, collecting sufficient data is costly/complex in industrial settings, and robustness when applying adaptive AI controllers during operation remains a point of concern.

TuPIC aims to address these issues: we aim to use data-driven AI methods but

• use them to tune parametrized controllers to facilitate validation and time/data/experiment efficiency
• tune only during commissioning, after which the parameters of the controllers will remain fixed, which guarantees safety and stability and simplifies the validation procedure
• increase the robustness and transferability to other tasks, conditions and systems

The methods from TuPIC will be validated on two main use cases: the control of a liquid cooling system and the control of a quadruped robot.