Data Exploration Using Bayesian Surrogates for Costly Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optimization methods for semiconductor design and process development are costly and inefficient due to the unknown nature of the objective function and the high cost of evaluating function values, making it difficult to derive an optimal solution quickly.
Innovation Solution
A data exploration apparatus and method using Bayesian optimization with a target-based acquisition function to iteratively refine input data based on a surrogate function, allowing for efficient optimization by gradually achieving set targets through repetitive evaluation and user interface interaction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If Bayesian optimization is used to reduce the number of evaluations, then the cost and time for optimization are reduced, but the objective function is unknown and expensive to evaluate
Solution Approach 1:
The patent introduces a surrogate function as an intermediary to represent the unknown objective function. This surrogate function is built from limited evaluation data and allows the optimization process to proceed without directly querying the expensive black-box objective function at every step, thus resolving the contradiction between reduced evaluation needs and unknown function characteristics
Solution Approach 2:
The patent implements a feedback mechanism where the surrogate function is continuously updated with new evaluation results, and the acquisition function uses this updated information to determine the next evaluation point. This iterative feedback loop allows the system to adapt to the unknown objective function while minimizing the number of expensive evaluations required
2Measurement precision
If the objective function is evaluated frequently to ensure accuracy, then the precision of optimization is improved, but the cost and time required increase significantly
Solution Approach 1:
The patent applies partial action by evaluating the objective function only at strategically selected points determined by the acquisition function, rather than exhaustively sampling the entire search space. This partial evaluation approach maintains optimization precision while dramatically improving evaluation efficiency by avoiding redundant or low-value queries
Solution Approach 2:
The patent changes the parameter representation by using the surrogate function's predicted mean and variance (from Gaussian process modeling) to guide evaluation selection. Instead of uniform or grid-based sampling, the system dynamically adjusts evaluation parameters based on the learned function characteristics, achieving high precision with fewer evaluations
3Productivity
If a surrogate function is used to model the objective function, then the number of evaluations is reduced, but the accuracy of the model may compromise the quality of the optimal solution
Solution Approach 1:
The patent uses feedback to continuously refine the surrogate function by incorporating new evaluation results into the Gaussian process model. This iterative updating ensures that the surrogate function progressively becomes more accurate and better represents the true objective function, maintaining solution quality while reducing evaluation needs
Solution Approach 2:
The patent replaces direct mechanical evaluation of the expensive objective function with a computational surrogate model (Gaussian process). This substitution allows for rapid, low-cost predictions that guide the optimization process, while periodic ground-truth evaluations ensure the surrogate remains accurate without requiring exhaustive sampling
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A processor-implemented method with data exploration includes: setting first input data and a first target condition; predicting first output data corresponding to the first input data using a first function that models an objective function; and determining second input data using a second function that provides a result of comparison between the first output data and the first target condition.