Machine Learning MPC Training for Real-Time Control Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional model-predictive control methods face challenges in solving optimization problems in real-time due to insufficient iterations, leading to suboptimal control solutions and uncertainty in system behavior, especially in safety-critical applications.
Innovation Solution
The use of machine learning methods to approximate the solution of the optimization problem for model-predictive control, specifically by training a machine learning system to evaluate the quality of states and select control signals, while incorporating slack variables to ensure stability and meet secondary conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional iterative optimization methods are used to solve the optimization problem in model-predictive control, then the control solution can be obtained with sufficient iterations, but the computation time exceeds real-time requirements
Solution Approach 1:
The patent applies preliminary action by pre-training a machine learning system offline to learn the mapping between system states and optimal control solutions. During real-time operation, the pre-trained model directly predicts control signals without requiring iterative optimization, thus achieving both high accuracy and real-time performance. The offline training phase performs the computationally intensive work in advance.
Solution Approach 2:
The patent substitutes the mechanical iterative optimization process with a machine learning-based prediction system. Instead of repeatedly solving optimization equations through computational iterations, the system uses a trained neural network or learning model to directly predict control actions, replacing the traditional computational mechanics with a learned approximation that runs much faster.
2Productivity
If the number of iterations is reduced to meet real-time requirements, then computation time is reduced, but the optimization solution becomes suboptimal
Solution Approach 1:
By performing the optimization learning in advance during an offline training phase, the system accumulates sufficient iteration results beforehand. The pre-trained model encapsulates the optimal solutions from extensive offline computations, enabling real-time deployment with high solution quality without requiring repeated iterations during runtime.
Solution Approach 2:
The patent creates a copy of the optimization knowledge by training a machine learning model to replicate the behavior of the iterative optimization process. The model learns to approximate the optimal control policy from training data generated by thorough offline optimization, allowing it to reproduce high-quality solutions rapidly during real-time operation without performing the full iterative process again.
3Productivity
If a suboptimal control solution is applied to the system, then real-time control is maintained, but the system behavior becomes unpredictable especially in safety-critical applications
Solution Approach 1:
The system performs comprehensive training in advance to ensure the model learns reliable and predictable control behaviors. During offline training, the model is exposed to diverse system states and optimal control actions, enabling it to generalize reliably to unseen situations during real-time operation, thus maintaining both speed and predictability.
Solution Approach 2:
The patent incorporates feedback mechanisms during the training phase where the model learns from the consequences of control actions through reinforcement learning or supervised learning with labeled optimal solutions. This feedback ensures the model learns not only to act quickly but also to predict system behavior accurately, improving reliability while maintaining real-time performance.
Data Source
AI summary
A computer-implemented method for training a machine learning system a model-predictive control of a technical system. The machine learning system is configured, with respect to an operating state and/or an environmental state of the technical system to be controlled, to ascertain a value that characterizes a quality of the state. The method includes ascertaining a plurality of operating states and/or environmental states of the technical system to be controlled; ascertaining a quality value of a state of the plurality of operating states and/or environmental states, wherein the quality value characterizes a quality of the state with respect to a model-predictive control; and training the machine learning system through supervised training of the machine learning system, wherein the state is used as input of the machine learning system and the ascertained quality value is used as the desired output of the machine learning system.