Neural Control Model Inversion With Regularized Signal Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Model predictive control (MPC) methods in robotics and automation face challenges due to the complexity and noise of automatically generated machine learning models, which can produce unsuitable control signals by optimizing artefacts rather than real-world system dynamics.
Innovation Solution
A method involving a neural network with a first component trained to predict the target system's state and a second component trained as a regularizer to invert the first component, using optimization with a regularization term to generate control signals, potentially implemented with denoising, generative adversarial, or variational autoencoder networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If machine learning methods are used to automatically generate a model of the target system, then the model can be constructed without requiring a sufficiently accurate physical model and can be automatically updated, but the generated model becomes very complex and internally noisy, producing artefacts that do not correspond to real-world system dynamics
Solution Approach 1:
The patent segments the neural network model into two distinct components: a forward model that predicts system states and a regularizer that enforces physical consistency. This segmentation allows each component to specialize - the forward model captures complex patterns from data while the regularizer ensures predictions align with physical laws, thereby resolving the contradiction between ease of construction and reliability
Solution Approach 2:
The regularizer acts as an intermediary between the data-driven forward model and the physical reality. It mediates the predictions by filtering out artefacts and ensuring that control actions correspond to real-world dynamics, thus maintaining reliability while preserving the ease of automatic model generation
2Extent of automation
If control actions are planned by optimizing the outcome predicted by the automatically generated model, then control signals can be generated automatically, but the optimization finds points in input space which are artefacts of the model and do not correspond to dynamics of the real world system
Solution Approach 1:
The regularizer serves as an intermediary in the optimization process, ensuring that automatically generated control actions are filtered through physical consistency constraints. This maintains automation while preventing artefact-driven control signals, resolving the contradiction between automation extent and control validity
Solution Approach 2:
The regularizer provides feedback during the optimization process by penalizing predictions that violate physical constraints. This feedback mechanism guides the optimization toward control actions that are both automated and physically valid, resolving the contradiction between automation and reliability
Data Source
AI summary
The invention relates to a method for generating a control system for a target system, wherein: operational data is received; a first neural model component is trained with the received operational data for generating a prediction on a state of the target system based on the received operational data; a second neural model component is trained with the operational data for generating a regularizer for use in inverting the first neural model component; and the control system is generated by inverting the first neural model component by optimization and arranging to apply the regularizer generated with the second neural model component in the optimization. The invention relates also to a system and a computer program product.