Meta-Learned Evolution Strategy Classifier for Black Box Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing evolution strategies for black box optimization lack the ability to incorporate prior knowledge about the structure of the optimization problem, leading to suboptimal performance, as they rely on predefined heuristics and hand-tuned hyperparameters without utilizing knowledge about the function being optimized.
Innovation Solution
The method involves meta-learning a black box optimization algorithm using machine learning techniques, such as deep learning, to convert static hyperparameters into learnable parameters and update evolution strategy parameters using meta-learning parameters trained with gradient descent, allowing the algorithm to improve optimization performance on specific classes of functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If evolution strategies use predefined heuristics and hand-tuned hyperparameters, then the algorithm can operate without prior knowledge of the function structure, but the optimization performance becomes suboptimal
Solution Approach 1:
The patent applies preliminary action by training the evolution strategy algorithm on a set of training functions before actual optimization. The algorithm learns function-specific parameters and structures in advance through meta-learning, enabling it to adapt quickly to new functions without requiring extensive hand-tuning during the actual optimization process
Solution Approach 2:
The patent changes parameters by converting static, hand-tuned hyperparameters into dynamic, learnable parameters. The algorithm learns optimal parameter values and function-specific characteristics from training data, allowing parameters to adapt based on the function being optimized rather than remaining fixed across all optimization tasks
2Device complexity
If evolution strategies rely on hand-tuned hyperparameters, then the algorithm structure remains simple and interpretable, but the convergence speed and efficiency are reduced
Solution Approach 1:
The patent performs preliminary learning on a dataset of training functions to pre-compute optimal hyperparameters and function-specific parameters. This preliminary action reduces the convergence time during actual optimization by having the algorithm already adapted to the function class it will optimize, without requiring complex real-time adjustments
Solution Approach 2:
The algorithm serves itself by automatically learning optimal parameters from training data without requiring manual intervention. The meta-learning process enables the algorithm to self-adjust its hyperparameters and internal structures based on patterns learned from similar functions, reducing both time loss and human effort
3Ease of operation
If evolution strategies use static hyperparameters, then the algorithm is easier to implement and control, but it cannot leverage prior knowledge about function structure
Solution Approach 1:
The patent transforms static hyperparameters into dynamic, learnable parameters that can capture function structure knowledge. The algorithm learns parameters specific to different function classes from training data, enabling it to utilize prior knowledge about function structures while maintaining a relatively simple implementation framework
Solution Approach 2:
The patent introduces an intermediary meta-learning layer that bridges the gap between simple algorithm structure and complex function knowledge. This intermediary layer learns from training functions and provides function-specific adaptations to the base evolution strategy, allowing the algorithm to leverage prior knowledge without fundamentally complicating the core optimization process
Data Source
AI summary
A computational method for training a meta-learned, evolution strategy black box optimization classifier. The method includes receiving one or more training functions and one or more initial meta-learning parameters of the optimization classifier. The method further includes sampling a sampled objective function from the one or more training functions and an initial mean of the sampled function. The method also includes computing a set of T number of means by running the meta learned, evolution strategy classifier on the sampled objective function using the initial mean for T number of steps in t=1, . . . , T. The method also includes computing a loss function from the set of T number of means. The method further includes updating the one or more initial meta-learning parameters of the optimization classifier in response to a characteristic of the loss function.