Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimization of Actuarial Neural Networks with Response Surface Methodology

Belguutei Ariuntugs, Kehelwala Dewage Gayan Madurang

TL;DR

This study utilizes a factorial design and response surface methodology (RSM) to optimize CANN performance, effectively explores the hyperparameter space and captures potential curvature, outperforming traditional grid search.

Abstract

In the data-driven world of actuarial science, machine learning (ML) plays a crucial role in predictive modeling, enhancing risk assessment and pricing strategies. Neural networks, specifically combined actuarial neural networks (CANN), are vital for tasks such as mortality forecasting and pricing. However, optimizing hyperparameters (e.g., learning rates, layers) is essential for resource efficiency. This study utilizes a factorial design and response surface methodology (RSM) to optimize CANN performance. RSM effectively explores the hyperparameter space and captures potential curvature, outperforming traditional grid search. Our results show accurate performance predictions, identifying critical hyperparameters. By dropping statistically insignificant hyperparameters, we reduced runs from 288 to 188, with negligible loss in accuracy, achieving near-optimal out-of-sample Poisson deviance loss.

Optimization of Actuarial Neural Networks with Response Surface Methodology

TL;DR

This study utilizes a factorial design and response surface methodology (RSM) to optimize CANN performance, effectively explores the hyperparameter space and captures potential curvature, outperforming traditional grid search.

Abstract

In the data-driven world of actuarial science, machine learning (ML) plays a crucial role in predictive modeling, enhancing risk assessment and pricing strategies. Neural networks, specifically combined actuarial neural networks (CANN), are vital for tasks such as mortality forecasting and pricing. However, optimizing hyperparameters (e.g., learning rates, layers) is essential for resource efficiency. This study utilizes a factorial design and response surface methodology (RSM) to optimize CANN performance. RSM effectively explores the hyperparameter space and captures potential curvature, outperforming traditional grid search. Our results show accurate performance predictions, identifying critical hyperparameters. By dropping statistically insignificant hyperparameters, we reduced runs from 288 to 188, with negligible loss in accuracy, achieving near-optimal out-of-sample Poisson deviance loss.

Paper Structure

This paper contains 17 sections, 15 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Response Surface Methodology Based Hyperparameter Tuning
  3. Related Works
  4. Model Description and Data Set
  5. Combined Actuarial Linear Model
  6. CANN Model Assumptions
  7. French Motor Third Party Liability Insurance Data
  8. Data Processing for GLM
  9. Data Processing for FFNN
  10. Computational Experiments and Setup
  11. CANN Implementation
  12. Initial Experiments
  13. Path of Steepest Descent
  14. Second Order Experimental Design Setup
  15. Second Order Experimental Results and Analysis
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Two Factor Response Surface Design
  • Figure 2: Neural Network architecture with $Lr=3$ hidden layers (20, 15, 10 neurons respectively) and GLM with skip connection
  • Figure 3: freMTPL2freq Summary