Multi-Parameter Circuit Design Optimization via Surrogate Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current model-building optimization algorithms face challenges in handling a large number of design variables, particularly in multi-parameter design optimization for analog, mixed-signal, and custom digital circuit designs, due to limitations in balancing exploration and exploitation, scalability issues with regression models, and inefficient inner optimization algorithms.
Innovation Solution
The method involves generating a set of candidate designs, calculating performance metric values with associated uncertainties, building surrogate models, performing multi-objective optimization to maximize uncertainty, and iteratively refining the design set until predetermined criteria are met, using an ensemble-style framework that allows for various regressors and stochastic optimization techniques to balance exploration and exploitation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional model-building optimization algorithms are used with a large number of design variables, then the optimization can handle complex multi-parameter designs, but the algorithms suffer from scalability issues and inability to effectively balance exploration and exploitation
Solution Approach 1:
The optimization algorithm segments the design space into multiple regions and uses surrogate models for each region. This allows the algorithm to handle complex multi-parameter designs by dividing the overall optimization problem into smaller, more manageable sub-problems, thereby improving scalability and reducing the complexity of the overall algorithm.
Solution Approach 2:
The patent introduces surrogate models as intermediary components between the actual performance metrics and the optimization process. These surrogate models approximate the complex relationships between design variables and performance metrics, enabling the algorithm to navigate complex design spaces more efficiently without directly computing all performance metrics, thus improving both adaptability and scalability.
2Productivity
If regression models are used to approximate performance metrics, then the optimization process becomes faster, but the models face scalability issues with a large number of design variables
Solution Approach 1:
The patent divides the design space into multiple regions and trains separate surrogate models for each region. This segmentation allows each individual regression model to remain computationally lightweight and scalable, while the collective system of models can handle a large number of design variables effectively. Each local model only needs to approximate relationships within its specific region, reducing the complexity burden on any single model.
Solution Approach 2:
The patent transforms the optimization problem by adding an additional dimension of regional classification. Instead of training a single monolithic regression model that must handle all design variables simultaneously, the system creates multiple simpler models organized by design space regions. This dimensional approach to modeling improves scalability while maintaining the speed benefits of regression-based approximation.
3Reliability
If the optimization algorithm maximizes uncertainty of surrogate models, then exploration of design space is improved, but the balance between exploration and exploitation becomes more difficult to achieve
Solution Approach 1:
The patent segments the optimization strategy into distinct phases: an exploration phase where surrogate models with maximized uncertainty are selected to identify promising regions, and an exploitation phase where models with minimized uncertainty are used to refine solutions. This segmentation resolves the complexity of continuously balancing exploration and exploitation by handling each phase separately with dedicated strategies, thereby improving reliability while managing computational complexity.
Solution Approach 2:
The patent implements a dynamic switching mechanism between exploration and exploitation modes based on the current state of the optimization process and the characteristics of the design space. The algorithm adapts its behavior by selecting surrogate models with different uncertainty characteristics at different stages, creating a dynamic optimization strategy that automatically balances exploration and exploitation without requiring complex manual control, thus improving reliability while reducing strategic complexity.
Data Source
AI summary
A method and system for performing multi-objective optimization of a multi-parameter design having several variables and performance metrics. The optimization objectives include the performance values of surrogate models of the performance metrics and the uncertainty in the surrogate models. The uncertainty is always maximized while the performance metrics can be maximized or minimized in accordance with the definitions of the respective performance metrics. Alternatively, one of the optimization objectives can be the value of a user-defined cost function of the multi-parameter design, the cost function depending from the performance metrics and/or the variables. In this case, the other objective is the uncertainty of the cost function, which is maximized. The multi-parameter designs include electrical circuit designs such as analog, mixed-signal, and custom digital circuits.