Parameter Search Automation Efficiency via Bayesian Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multi-task Bayesian optimization methods are inefficient when the correlation between target and reference observation data is low, leading to incorrect model estimation and reduced parameter search efficiency.
Innovation Solution
A parameter-searching method that estimates the posterior distribution of a model function by correcting for the variation between target and reference observation data within the parameter space, using methods such as shift estimation to improve correlation and accuracy, allowing for efficient parameter value determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional multi-task Bayesian optimization is used to search for optimal parameters, then the parameter search can be automated, but the search efficiency deteriorates when correlation between target and reference observation data is low
Solution Approach 1:
The patent introduces a correlation coefficient calculation as an intermediary step to assess the relationship between target and reference observation data. When the correlation coefficient falls below a threshold, the system switches to single-task Bayesian optimization instead of forcing multi-task optimization, thereby avoiding inefficient searches while maintaining automation.
Solution Approach 2:
The patent dynamically adjusts the optimization approach based on the calculated correlation coefficient. The system transitions between multi-task and single-task Bayesian optimization modes depending on the data correlation level, optimizing search efficiency adaptively rather than using a fixed approach.
2Adaptability or versatility
If multi-task Bayesian optimization is applied, then parameter search can leverage reference data, but model estimation accuracy deteriorates when data correlation is low
Solution Approach 1:
The correlation coefficient serves as an intermediary indicator to evaluate whether reference observation data should be used for model estimation. By calculating this coefficient first, the system determines the appropriateness of leveraging reference data, ensuring accuracy is maintained while preserving adaptability to use available data when appropriate.
3Reliability
If parameter search is performed with many measurement repetitions, then optimal parameter value can be found with higher confidence, but time consumption and sample consumption increase
Solution Approach 1:
The patent implements feedback through correlation coefficient calculation to guide the parameter search process. By assessing data correlation before initiating optimization, the system avoids unnecessary repeated measurements that would consume time and samples, while still achieving reliable optimal parameter identification through the appropriate optimization method selection.
Data Source
AI summary
A model estimator (11) estimates a model function of an analyzing system (20) based on target observation data and reference observation data. For this estimation, the amount of variation between the model of the target system and that of a reference model is estimated from the target observation data and reference observation data. The model function of the target model is estimated after the observation data are corrected based on the estimated amount of variation. A parameter determiner (12) calculates an acquisition function based on the mean and covariance of the model function, and determines a parameter value for the next observation, using the acquisition function. A data acquirer (13) sets the parameter value in the analyzing system (20) and acquires a corresponding observed value. A loop process with the feedback of the observation data is repeated to determine an optimal parameter value. In the case of a parameter search using a multi-task Bayesian optimization method, the present technique improves the efficiency of the parameter search even if the correlation between the target observation data and reference observation data is low.