[Paper Review] Neuroevolution in Deep Neural Networks: Current Trends and Future Challenges

Edgar Galván, Peter Mooney|arXiv (Cornell University)|Jun 9, 2020

Evolutionary Algorithms and Applications170 references172 citations

TL;DR

This paper presents a comprehensive survey of neuroevolution in deep neural networks (DNNs), evaluating evolutionary algorithms (EAs) for automated architecture search and training. It identifies key trends, challenges, and future research directions, highlighting EAs as a computationally efficient alternative to reinforcement learning with strong potential for improving DNN performance and automation in real-world applications.

ABSTRACT

A variety of methods have been applied to the architectural configuration and learning or training of artificial deep neural networks (DNN). These methods play a crucial role in the success or failure of the DNN for most problems and applications. Evolutionary Algorithms (EAs) are gaining momentum as a computationally feasible method for the automated optimisation and training of DNNs. Neuroevolution is a term which describes these processes of automated configuration and training of DNNs using EAs. While many works exist in the literature, no comprehensive surveys currently exist focusing exclusively on the strengths and limitations of using neuroevolution approaches in DNNs. Prolonged absence of such surveys can lead to a disjointed and fragmented field preventing DNNs researchers potentially adopting neuroevolutionary methods in their own research, resulting in lost opportunities for improving performance and wider application within real-world deep learning problems. This paper presents a comprehensive survey, discussion and evaluation of the state-of-the-art works on using EAs for architectural configuration and training of DNNs. Based on this survey, the paper highlights the most pertinent current issues and challenges in neuroevolution and identifies multiple promising future research directions.

Motivation & Objective

• To address the fragmented state of neuroevolution research by providing a unified, comprehensive survey focused exclusively on evolutionary algorithms (EAs) in deep neural network (DNN) architecture configuration and training.
• To identify and analyze the strengths, limitations, and current challenges of EA-based neuroevolution in DNNs, particularly in comparison to reinforcement learning and other hyperparameter optimization methods.
• To highlight the computational advantages of neuroevolution, such as reduced training time compared to reinforcement learning, and its potential for broader adoption in real-world deep learning applications.
• To advocate for standardized benchmarking and reproducible studies in neuroevolution to enable fair comparison and progress in the field.
• To identify promising future research directions, including improved genetic operators, standardized evaluation protocols, and expansion into new application domains beyond image classification.

Proposed method

• Conducted an extensive literature review across multiple databases (arXiv, IEEE Xplore, ACM, Google Scholar, etc.) to identify and analyze state-of-the-art works on EA-based DNN architecture search and training.
• Classified and evaluated EA-based approaches based on their representation (e.g., weight-based, architecture-based), evolutionary paradigms (e.g., genetic algorithms, evolution strategies, neuroevolution of augmenting topologies), and training strategies.
• Reviewed and compared various deep learning architectures used in neuroevolution, including Convolutional Neural Networks (CNNs), Autoencoders (AEs), Deep Belief Networks (DBNs), Restricted Boltzmann Machines (RBMs), and Recurrent Neural Networks (RNNs).
• Analyzed the role of different EA components—such as selection, mutation, and crossover—on the performance and convergence of neuroevolution in DNNs, emphasizing the need for well-designed genetic operators.
• Evaluated benchmark datasets (e.g., MNIST, CIFAR-10, ILSVRC2012) commonly used in neuroevolution studies to assess reproducibility and standardization.
• Discussed the importance of open, publicly available datasets and the challenges posed by proprietary or non-standardized data in comparing results across studies.

Experimental results

Research questions

• RQ1 What are the current trends and dominant methodologies in using evolutionary algorithms for architecture search and training in deep neural networks?
• RQ2 How do EA-based neuroevolution approaches compare in performance and computational efficiency to reinforcement learning-based methods in DNN architecture search?
• RQ3 What are the key challenges and limitations in applying neuroevolution to DNNs, particularly regarding representation, scalability, and reproducibility?
• RQ4 Why is there a lack of standardized benchmarking in neuroevolution research, and how does this hinder progress and fair comparison across studies?
• RQ5 What future research directions are most promising for advancing neuroevolution in deep learning, especially in expanding beyond image classification tasks?

Key findings

• Neuroevolution using EAs has shown competitive performance in DNN architecture search and training, often requiring less computational time than reinforcement learning methods.
• Evolutionary algorithms are particularly effective in handling complex, non-linear, and discontinuous search spaces common in DNN hyperparameter and architecture optimization.
• Despite growing interest, the field lacks standardized benchmarking protocols and publicly available datasets, leading to inconsistent comparisons and reduced reproducibility across studies.
• The paper identifies a critical need for standardized scientific studies in neuroevolution to isolate the impact of specific EA components (e.g., representation, operators, learning type) on performance.
• Commonly used datasets such as MNIST, CIFAR-10, and ILSVRC2012 are widely adopted in neuroevolution research, but their limited scope restricts evaluation to vision-centric tasks.
• Future research should prioritize the development of domain-specific benchmarks beyond image classification and the creation of open, shared datasets to enable fair and reproducible evaluation of neuroevolution techniques.