Fitness Approximation through Machine Learning - Submitted to Special Issue on Machine Learning Assisted Evolutionary Computation

Aug 3, 2026

Upload by Levi

A novel fitness-approximation approach for genetic algorithms uses machine-learning surrogate models that adapt to the evolutionary process state. The method keeps a dataset of sampled individuals with their true fitness scores, updating the surrogate continuously during a run. Multiple design choices are compared, including switching between approximate and actual fitness, how to sample individuals for the dataset, and how to weight samples. Experiments in Gymnasium-based game environments show improved evolutionary runtimes and fitness outcomes close to, or slightly below, full true-fitness GAs, with computation savings when fitness evaluation is costly.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Fitness approximation with machine learning

Evolution control via dynamic switching

Experimental evaluation on Gymnasium games

Fitness Approximation through Machine Learning Submitted to Special Issue on Machine Learning Assisted Evolutionary Computation Itai Tzruia , Tomer Halperin , Moshe Sipper , Achiya Elyasaf arXiv :2309 .03318v2 [ cs .NE] 21 May 2024 Abstract—We present a novel approach to performing fitness approximation in genetic algorithms (GAs) using machinelearning (ML) models, through dynamic adaptation to the evolutionary state. Maintaining a dataset of sampled individuals along with their actual fitness scores, we continually update a fitness-approximation ML model throughout an evolutionary run. We compare different methods for: 1) switching between actual and approximate fitness, 2) sampling the population, and 3) weighting the samples. Experimental findings demonstrate significant improvement in evolutionary runtimes, with fitness scores that are either identical or slightly lower than that of the fully run GA—depending on the ratio of approximate-to-actualfitness computation. Although we focus on evolutionary agents in Gymnasium (game) simulators—where fitness computation is costly—our approach is generic and can be easily applied to many different domains. Index Terms—genetic algorithm, machine learning, fitness approximation, surrogate models, regression, agent simulation. I. INTRODUCTION A genetic algorithm (GA) is a population-based metaheuristic optimization algorithm that operates on a population of candidate solutions, referred to as individuals, iteratively improving the quality of solutions over generations. GAs employselection, crossover, and mutation operators to generate new individuals based on their fitness values, computed using a fitness function [19] . GAs have been widely used for solving optimization problems in various domains, such as telecommunication systems [21], energy systems [31], and medicine [17] . Further, GAs can be used to evolve agents in game simulators. For example, Garc´ıa-Snchez et al. [12] employed a GA to enhance agent strategies in Hearthstone, a popular collectible card game, and Elyasaf et al. [11] evolved top-notch solvers for the game of FreeCell. An accurate evaluation of a fitness function is often computationally expensive, particularly in complex and highdimensional domains, such as games. In fact, a GA spends most of its time computing fitness. To mitigate this cost, fitness approximation techniques have been proposed to estimate the fitness values of individuals based on a set of features or characteristics. This paper focuses on performing fitness approximation in genetic algorithms A. Elyasaf is with the Software and Information Systems Engineering Department, Ben-Gurion University of the Negev, Israel. E-mail: achiya@bgu .ac .il I. Tzruia, T. Halperin, and M. Sipper are with the Computer Science Department, Ben-Gurion University of the Negev, Israel. E-mail: [itaitz@post.bgu.ac.il](itaitz@post.bgu.ac.il), [tomerhal@post.bgu.ac.il](tomerhal@post.bgu.ac.il), and [sipper@bgu.ac.il](sipper@bgu.ac.il) using machine learning (ML) models. Specifically, we propose to maintain a dataset of individuals and their actual fitness values, and to learn a surrogate fitness-approximation model based on this dataset. Relying only on approximate fitness scores might cause the GA to converge to a false optimum. To address this, the evolutionary process can be controlled by combining approximate and actual fitness evaluations. This process is referred to as evolution control [23] . While there are static evolution-control methods [6, 36, 37], our method dynamically switches between true and approximate fitness evaluations, adapting to the evolutionary process’s current state. We analyze several options for: 1) switch conditions between using the actual fitness function and the approximate one, 2) sampling the search space for creating the dataset, and 3) weighting the samples in the dataset. We evaluate our approach on three games implemented by Gymnasium, a framew

FAQs

What problem does the paper address in genetic algorithms?

Evaluating fitness functions is often computationally expensive, so the paper targets reducing this cost while maintaining optimization quality.

How does the proposed approach approximate fitness during an evolutionary run?

It maintains a dataset of individuals and their actual fitness values, then trains and updates a surrogate fitness-approximation model throughout the run.

What comparisons does the paper make to validate the method?

It compares switching between actual and approximate fitness, population sampling strategies, and sample weighting schemes.

Aug 3, 2026

Upload by Levi