Control Strategy Parameter Tuning With Bayesian Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional model-based reinforcement learning methods for control systems face inefficiencies due to unsuitable model structures, leading to suboptimal parameter adjustments and high interaction times with the environment, while model-free methods require excessive interaction, and Bayesian optimization with large parameter sets leads to complex and time-consuming optimization processes.

Innovation Solution

A method using Bayesian optimization with a trainable regression function to iteratively optimize control model parameters, defining initial model parameter domains based on dynamic system models and maximizing a posteriori estimates, balancing exploration and exploitation to minimize evaluations, and employing a Gaussian process regression to approximate system behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If model-free reinforcement learning methods are used, then no knowledge of the environment is required, but the system's interaction time with the environment during learning is very high

Engineering Contradiction:
Improveenvironment independenceVSAvoidinteraction time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent creates initial control models based on dynamic system models before the Bayesian optimization process begins. These preliminary models provide a starting point that reduces the exploration needed during optimization, thereby decreasing interaction time while maintaining environment independence through the use of system models

Inventive Principle:
Principle #10Preliminary action

2Productivity

If conventional model-based methods are used with a defined model structure, then parameter adjustments are simple and efficient, but the chosen model structure may be unsuitable leading to suboptimal results

Engineering Contradiction:
Improveparameter optimization efficiencyVSAvoidcontrol performance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent employs Bayesian optimization to dynamically adjust model parameters based on performance feedback from the quality function. Instead of using a fixed model structure with static parameters, the optimization process adaptively tunes parameters iteratively, allowing the system to find optimal configurations that balance efficiency and performance reliability

Inventive Principle:
Principle #15Dynamics

3Reliability

If Bayesian optimization is applied with a large number of model parameters, then comprehensive control optimization is achieved, but the optimization process becomes very complex and requires long training times

Engineering Contradiction:
Improvecontrol optimization qualityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent segments the parameter optimization process by creating multiple initial control models with different parameter configurations. The Bayesian optimization then systematically evaluates and compares these segmented models, allowing comprehensive optimization to be achieved through structured comparison rather than exhaustive search, thereby reducing training time

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

By pre-defining a reasonable parameter domain based on dynamic system models before optimization begins, the patent reduces the search space that Bayesian optimization needs to explore. This preliminary constraint setting maintains comprehensive optimization quality while significantly reducing the computational time required

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3748556B1Method and device for determining a control strategy for a technical system
Publication Date: 2026.02.25 ROBERT BOSCH GMBH
  • EP3748556B1 patent drawingFigure 1
  • EP3748556B1 patent drawingFigure 2
  • EP3748556B1 patent drawingFigure 3

AI summary

The invention relates to a computer-implemented method for creating a control system for a technical system (2) using a Bayesian optimization method, wherein the control system is created and executable based on model parameters (θ) of a control model, and wherein the following steps are performed to optimize the control system: - providing a quality function whose form corresponds to a trainable regression function and which evaluates the quality of a control system for the technical system (2) based on model parameters (θ); - performing (S4) a Bayesian optimization method based on the quality function to iteratively optimize a set of model parameters (θ) within an initially specified model parameter domain (θ);wherein, prior to performing the Bayesian optimization procedure, initial control models are created based on several dynamic system models and the model parameter domain (0), which specifies the permissible ranges of values ​​for the model parameters (θ), is determined by determining a distribution of the model parameters (θ) of the initial control models and, depending on an associated maximum a posteriori estimate (θMAP) and its respective standard deviation of the distribution of the model parameters (θ), the ranges of values ​​of the model parameters (θ) are initially defined.