Hyperparameter Tuning via Pareto Frontier Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models often employ suboptimal hyperparameters, leading to inefficient performance, and current optimization systems require complex coding and understanding of software and hardware components, making it difficult to effectively optimize these models.
Innovation Solution
A method and API for tuning hyperparameters using a multi-criteria approach that defines a joint tuning function, identifies a Pareto efficient frontier curve, applies metric thresholds, and searches for optimal hyperparameter values within a feasible section of this curve, enabling efficient optimization of competing performance objectives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple objective functions are optimized simultaneously, then model performance across multiple metrics is improved, but the optimization complexity increases due to divergent metric improvements
Solution Approach 1:
The patent introduces a joint tuning function as an intermediary that combines multiple objective functions into a single unified objective. This joint tuning function serves as a mediator that transforms the complex multi-objective optimization problem into a more manageable single-objective problem, allowing simultaneous optimization of multiple metrics without directly handling the divergent optimization relationships
Solution Approach 2:
The patent transforms the multi-objective optimization problem by changing the parameter representation from multiple separate objective functions to a single joint tuning function with modified parameters. This parameter transformation enables the system to optimize multiple metrics simultaneously while working within a simplified single-objective framework
2Reliability
If existing optimization systems are used to tune hyperparameters, then model performance can be improved, but the interface complexity requires significant coding capabilities and understanding of underlying components
Solution Approach 1:
The patent implements an automated hyperparameter optimization system that performs tuning operations without requiring users to have deep coding capabilities or understand underlying software and hardware components. The system serves itself by automatically defining joint tuning functions, executing tuning operations, and identifying optimal hyperparameter values, making the optimization process accessible to users with minimal technical expertise
3Ease of manufacture
If default hyperparameters are used, then model implementation is simplified, but predictive performance is suboptimal
Solution Approach 1:
The patent implements preliminary hyperparameter optimization by automatically tuning hyperparameters before model deployment. The system performs joint tuning operations and identifies optimal hyperparameter values in advance, ensuring that models are implemented with optimized parameters rather than defaults, thereby improving predictive performance while maintaining ease of implementation
Data Source
AI summary
Systems and methods for tuning hyperparameters of a model include receiving a tuning request for tuning hyperparameters, the tuning request includes a first and a second objective function for the machine learning model. The first and second objective functions may output metric values that do not improve uniformly. Systems and methods additionally include defining a joint tuning function that is based on a combination of the first and second objective functions; executing a tuning operation; identifying a Pareto efficient frontier curve defined by a plurality of distinct hyperparameter values; applying metric thresholds to the Pareto efficient frontier curve; demarcating the Pareto efficient frontier curve into at least a first infeasible section and a second feasible section; searching the second feasible section of the Pareto efficient frontier curve for one or more proposed hyperparameter values; and identifying at least a first set of proposed hyperparameter values based on the search.