Invertible Neural Networks for High-Dimensional Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for multi-objective black box optimization, such as gradient-free and gradient-based methods, face challenges in handling interdependent parameters, non-convex and noisy objective functions, and the need for diverse solutions, particularly in complex design problems like propulsion chain optimization for electric air vehicles and deep neural network architecture design.
Innovation Solution
The use of deep learning techniques involving invertible neural networks (INNs) and latent space learning to compress high-dimensional parameter spaces into lower-dimensional representations, allowing for the exploration of diverse solutions that satisfy multiple objectives through autoencoders and surrogate models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional gradient-free or gradient-based optimization methods are used for multi-objective black box optimization, then the optimization process can handle the complexity of interdependent parameters and non-convex noisy objective functions, but the computational cost and time consumption become excessively high
Solution Approach 1:
The patent applies preliminary action by pre-training an autoencoder to learn the latent representation of the parameter space before performing optimization. This pre-processing step compresses the high-dimensional parameter space into a lower-dimensional latent space, making subsequent optimization iterations faster while maintaining accuracy in handling interdependent parameters and non-convex noisy objective functions
Solution Approach 2:
The patent introduces an autoencoder as an intermediary between the high-dimensional parameter space and the optimization process. The autoencoder acts as a mediator that transforms complex parameter configurations into compressed latent representations, enabling efficient exploration of the solution space without sacrificing the ability to handle parameter interdependencies and noisy objectives
2Adaptability or versatility
If the parameter space is explored exhaustively to find diverse solutions satisfying multiple objectives, then solution diversity and optimality are improved, but the computational complexity and resource requirements increase significantly
Solution Approach 1:
The patent applies dimensionality change by transforming the optimization problem from the original high-dimensional parameter space into a lower-dimensional latent space using an autoencoder. This dimensional transformation allows for more efficient exploration of diverse solutions while reducing computational complexity, as the latent space captures the essential variations in parameter configurations with fewer dimensions
Solution Approach 2:
The patent uses the autoencoder to create compressed copies or representations of the parameter configurations in latent space. Instead of working with full high-dimensional parameter sets, the optimization process operates on these compressed latent representations, which preserve the essential information needed to evaluate multiple objectives and generate diverse solutions at reduced computational cost
Data Source
AI summary
In an example, a method of designing a system or architecture includes, receiving a plurality of parameter values and a set of requirements for a plurality of objective functions related to a design problem; compressing the plurality of parameters to generate a latent representation; forward processing, with one or more Invertible Neural Networks (INNs), the latent representation to generate a plurality of objective values corresponding to the plurality of the objective functions; inverse processing the plurality of objective values; and generating, based on the latent representation, a plurality of solutions to the design problem that satisfy the set of requirements for the plurality of objective functions.