Core Concepts
A machine learning-based optimization workflow is proposed to efficiently tune the numerical settings of boundary value problem solvers, reducing the time and domain expertise required in the process.
Abstract
The paper presents a machine learning-based optimization workflow for tuning the numerical settings of boundary value problem (BVP) solvers. The workflow consists of two main stages:
Machine Learning Process:
Feature Selection: Nine input features affecting the solvability and performance of the BVP solver are identified, including solver-specific settings and problem-specific parameters.
Data Collection: A dataset of 100,000 samples is generated for each of the ten reference BVP problems by sampling the input feature space.
Model Selection: Several binary classification and multi-output regression algorithms are evaluated, and the best-performing models are selected based on accuracy metrics.
Training and Testing: The machine learning models are trained and tested on the dataset, with the Random Forest algorithm performing the best for regression tasks.
Optimization Process:
Multi-Objective Optimization: The trained machine learning models are used to define a multi-objective optimization problem, aiming to minimize the residual error, number of ODE evaluations, and grid points required by the BVP solver.
Optimization Results: The Pareto front of optimal numerical settings is obtained using the Optuna optimization library, providing a set of trade-off solutions that balance the competing objectives.
The proposed workflow is evaluated for scalability, stability, and reliability. The results demonstrate the effectiveness of the machine learning-based approach in tuning the numerical settings of BVP solvers, reducing the time and domain expertise required compared to manual tuning.
Stats
The number of ODE evaluations required by the BVP solver can range from 0 to 600,000.
The number of grid points required by the BVP solver can range from 0 to 6,000.
The maximum residuum of the BVP solver's numerical solution can range from 0 to 0.001.
Quotes
"Several numerical differential equation solvers have been employed effectively over the years as an alternative to analytical solvers to quickly and conveniently solve differential equations."
"A systematic fine-tuning of these settings is required to obtain the desired solution and performance. Currently, these settings are either selected by trial and error or require domain expertise."
"In this paper, we propose a machine learning-based optimization workflow for fine-tuning the numerical settings to reduce the time and domain expertise required in the process."