Neural Networks for On-Chip Model Predictive Control: A Method to Build Optimized Training Datasets and its...

Area of Science:

• Control Systems Engineering
• Machine Learning
• Biomedical Engineering

Background:

• Neural networks (NNs) can emulate model predictive control (MPC) algorithms for efficient implementation on embedded devices.
• NN accuracy is highly dependent on training data composition, yet systematic optimization methods are underdeveloped.
• Existing methods for NN training data selection lack systematic optimization for replicating MPC behavior.

Purpose of the Study:

• Introduce and develop an efficient algorithm for generating optimally-sampled datasets (OSDs).
• To systematically optimize training data for neural networks mimicking model predictive control.
• Enhance the accuracy and efficiency of neural network-based control systems for embedded applications.

Main Methods:

• Defined optimally-sampled datasets (OSDs) as parametrized subsets of data preserving MPC information, avoiding duplicates, and achieving completeness.
• Developed an efficient algorithm for generating these OSDs.
• Trained NNs using OSDs to replicate a specific MPC algorithm for automated insulin delivery.

Main Results:

• Achieved a fourfold improvement in final accuracy for NNs trained with OSDs.
• Demonstrated the effectiveness of OSDs in replicating complex MPC behavior.
• Two OSD-trained NNs received regulatory clearance for clinical testing in automated insulin delivery.

Conclusions:

• OSDs provide a systematic and effective method for optimizing NN training data for MPC emulation.
• This approach significantly enhances NN performance on resource-constrained embedded platforms.
• The methodology paves the way for deploying advanced control algorithms in real-world applications, including medical devices.