[Paper Review] Investigating the effects Diversity Mechanisms have on Evolutionary Algorithms in Dynamic Environments

Matthew T. Hughes|arXiv (Cornell University)|Oct 9, 2016

Metaheuristic Optimization Algorithms Research19 references3 citations

TL;DR

This paper investigates how diversity mechanisms impact evolutionary algorithm performance in dynamic optimization problems, comparing techniques like fitness sharing, clearing, crowding, incest prevention, and island models. The key contribution is a novel hybrid algorithm that outperforms all others in tracking multiple moving peaks, achieving near-optimal fitness and maintaining population diversity across all dynamic test problems.

ABSTRACT

Evolutionary algorithms have been successfully applied to a variety of optimisation problems in stationary environments. However, many real world optimisation problems are set in dynamic environments where the success criteria shifts regularly. Population diversity affects algorithmic performance, particularly on multiobjective and dynamic problems. Diversity mechanisms are methods of altering evolutionary algorithms in a way that promotes the maintenance of population diversity. This project intends to measure and compare the performance effect a variety of diversity mechanisms have on an evolutionary algorithm when facing an assortment of dynamic problems.

Motivation & Objective

• To evaluate the impact of various diversity mechanisms on evolutionary algorithm performance in dynamic environments.
• To identify which mechanisms are most effective for maintaining population diversity during fitness landscape changes.
• To compare algorithmic performance across a range of dynamic benchmark problems, including moving peaks and height-changing landscapes.
• To develop and test a new hybrid algorithm combining strengths of existing diversity mechanisms for improved dynamic optimization.
• To assess the robustness and adaptability of diversity mechanisms under varying problem dynamics and parameter settings.

Proposed method

• Implemented a suite of evolutionary algorithms with distinct diversity mechanisms: fitness sharing, clearing, crowding, incest prevention, genotype removal, island models, and a novel hybrid algorithm.
• Applied these algorithms to six dynamic test problems: OneMax, TwoMax, One Moving Peak, Two Moving Peaks, Height Changing Peaks, and Moving Height Changing Peaks.
• Used DEAP (Distributed Evolutionary Algorithms in Python) for algorithm implementation and experimentation with standardized parameters.
• Measured performance using offline performance, maximum fitness, diversity metrics, and convergence stability across multiple runs.
• Adjusted bit-flip mutation probabilities to assess sensitivity of algorithms to dynamic changes in problem landscapes.
• Designed a new hybrid algorithm by integrating the strengths of multiple diversity mechanisms to enhance tracking of multiple moving peaks.

Experimental results

Research questions

• RQ1 How do different diversity mechanisms affect the ability of evolutionary algorithms to track multiple moving peaks in dynamic environments?
• RQ2 Which diversity mechanisms are most effective for maintaining population diversity during rapid fitness landscape changes?
• RQ3 Can a hybrid algorithm combining multiple diversity mechanisms outperform individual mechanisms on complex dynamic problems?
• RQ4 How does population diversity correlate with fitness performance and convergence stability in dynamic optimization tasks?
• RQ5 What is the impact of varying mutation rates and problem dynamics on the performance of diversity mechanisms?

Key findings

• The novel hybrid algorithm achieved the highest offline performance (94.910) on the One Moving Peak problem, outperforming all other algorithms.
• On the Two Moving Peaks problem, the new algorithm was the only one capable of consistently locating and maintaining individuals at both peaks, achieving 100% success in peak detection.
• For the Height Changing Peaks problem, the new algorithm maintained maximum fitness (100) even after peak movements, demonstrating superior stability and convergence.
• On the Moving Height Changing Peaks problem, the new algorithm achieved the highest offline performance (90.249) and maintained fitness near the global optimum despite dynamic changes.
• The new algorithm successfully tracked both peaks in all multi-peak problems, with 100% success rate in peak detection, while other algorithms failed to maintain subpopulations.
• Fitness sharing and clearing showed strong performance on single-peak problems but struggled with multi-peak tracking, while island models and incest prevention were less effective overall.