Neuroevolution in Deep Neural Networks: Current Trends and Future Challenges

• Architecture configuration and training strongly influence whether DNNs succeed or fail on applications.Source: §I · p. 1
• Architecture search includes evolution-based and reinforcement-learning methods, alongside random search, Bayesian optimisation, hill-climbing, grid search, and Monte Carlo approaches.Source: §I · p. 1
• Neuroevolution encompasses EA-based methods for automatically searching DNN architectures and training networks.Source: §I · p. 1
• The survey considers landmark and newer EAs across autoencoders, CNNs, DBNs, RBMs, RNNs, and LSTMs.Source: §I · p. 1
• Evolutionary Programming is identified as a landmark EA without known neuroevolution applications in the surveyed literature.Source: §I · p. 1
• The paper aims to provide a timely, comprehensive review of neuroevolution in DNNs and identify strengths, challenges, and future research directions.Source: §I · p. 2

Deep Neural Networks

• Deep learning transforms raw data into progressively more abstract and composite representations through multiple layers.Source: A. Deep Neural Networks · p. 2

Convolutional Neural Networks

• CNNs process grid-like data using local receptive fields, convolutional filters, and feature maps.Source: 1) Deep Learning Architecture:… · p. 2
• CNN efficiency relies on sparse interaction, parameter sharing, and equivariant representations.Source: 1) Deep Learning Architecture:… · p. 2
• CNN architectures have expanded from early LeNet designs to deeper models such as AlexNet, with rectified linear units becoming standard in neuroevolutionary CNN design.Source: 1) Deep Learning Architecture:… · p. 3

Autoencoders

• Autoencoders combine encoder and decoder functions to reconstruct inputs while preserving information in an alternative representation.Source: 2) Deep Learning Architecture:… · p. 3

Deep Belief Networks

• DBNs support feature extraction and can be formed by stacking RBMs, but their many parameters require substantial training time.Source: 3) Deep Learning Architecture: Deep… · p. 3

Recurrent Neural Networks

• RNNs process sequences through recurrent connections and internal state, supporting temporal behaviour and sequential prediction.Source: 4) Deep Learning Architecture: Other… · p. 3
• RNN training is difficult because of exploding and vanishing gradients, and practical look-back may cover only a few steps.Source: 4) Deep Learning Architecture: Other… · p. 3

Restricted Boltzmann Machines

• RBMs are energy-based networks whose hidden-unit activities can provide training data for higher-level RBMs.Source: 4) Deep Learning Architecture: Other… · p. 3
• Stacking RBMs lets networks learn features from features toward higher-level representations.Source: 4) Deep Learning Architecture: Other… · p. 4

Long Short Term Memory

• LSTMs are recurrent networks designed to learn long-term dependencies using memory blocks, cells, and adaptive gating units.Source: 4) Deep Learning Architecture: Other… · p. 4

Evolutionary Algorithms

• EAs evolve populations of encoded candidate solutions across generations using fitness evaluation, selection, crossover, and mutation.Source: B. Evolutionary Algorithms · p. 4

Genetic Algorithms

• Genetic algorithms are search methods that seek near-optimal solutions and have influenced machine learning and neural-network research.Source: 1) Evolutionary Algorithm: Genetic… · p. 4

Genetic Programming

• Genetic programming is automated programming that commonly represents individuals as tree-like structures generated from functional and terminal sets.Source: 2) Evolutionary Algorithm: Genetic… · p. 4

Common EA Process

• A common EA network-design process takes a dataset, generation count, population size, mutation probability, and crossover probability as inputs.Source: 2) Evolutionary Algorithm: Genetic… · p. 5
• Each generation applies selection, crossover, and mutation before evaluating the fitness of newly produced individuals.Source: 2) Evolutionary Algorithm: Genetic… · p. 5
• The final generation is returned as a set of evolved individuals with their fitness values.Source: 2) Evolutionary Algorithm: Genetic… · p. 5

• ES were introduced in the 1960s by Rechenberg and Schwefel for real-valued optimisation problems.Source: 3) Evolutionary Algorithm: Evolution… · p. 5
• Mutation is the main ES operator, while crossover is secondary and optional.Source: 3) Evolutionary Algorithm: Evolution… · p. 5
• The two historical ES forms are (µ, λ)-ES and (µ + λ)-ES, distinguished by how parents and offspring participate in selection.Source: 3) Evolutionary Algorithm: Evolution… · p. 5
• In ES notation, µ denotes the parent-population size and λ denotes the number of offspring produced before selection.Source: 3) Evolutionary Algorithm: Evolution… · p. 5
• Evolutionary Programming differs from ES mainly because it does not use crossover and normally produces M offspring from M parents.Source: 4) Evolutionary Algorithm · p. 5

• Differential Evolution (DE) is reported as efficient in continuous search spaces, robust, and fast-converging relative to other optimisation methods.Source: 5) Evolutionary Algorithm: Others · p. 5
• DE variants perturb population members using scaled differences between randomly selected, distinct population members.Source: 5) Evolutionary Algorithm: Others · p. 5
• Grammatical Evolution (GE) uses binary strings to select production rules in a Backus-Naur form grammar through genotype–phenotype mapping.Source: 5) Evolutionary Algorithm: Others · p. 5
• GE is presented as a special form of Genetic Programming in which evolution operates on representations rather than directly on programs.Source: 5) Evolutionary Algorithm: Others · p. 5
• NEAT evolves neural networks using historical marking, speciation, and incremental growth from a minimal starting structure.Source: 5) Evolutionary Algorithm: Others · p. 5
• EAs are gaining momentum in architecture search because they are gradient-free, population-based, and parallelisable, but their fitness evaluations can be computationally expensive.Source: A. The Motivation · p. 5, B. The Critique · p. 5
• With 20 individuals and 50,000 training samples, one EA generation requires one million fitness-function evaluations.Source: B. The Critique · p. 5

• CNN architectures have been evolved using tree-based GP, genetic algorithms, grammatical evolution, Cartesian GP, graph encodings, and constrained binary representations.Source: C. Deep Learning Architecture:… · pp. 6–8
• NASNet uses repeated Normal Cells and Reduction Cells, where Reduction Cells halve feature-map height and width while same-type cells share an architecture.Source: C. Deep Learning Architecture:… · p. 6, Figure 1 · p. 7
• NASNet-inspired mutations select a random cell and pairwise combination before applying hidden-state or operation mutations.Source: C. Deep Learning Architecture:… · p. 6
• Real et al. encoded architectures as graphs and found that three crossover forms did not improve mutation-based results.Source: C. Deep Learning Architecture:… · pp. 6–7
• Cartesian GP encodes CNN nodes and connections in a directed acyclic grid, with Suganuma et al. using a (1+λ) ES because individual evaluation was expensive.Source: C. Deep Learning Architecture:… · p. 7, Figure 2 · p. 8
• Hybrid approaches combined evolutionary architecture or representation search with backpropagation for weight optimisation.Source: C. Deep Learning Architecture:… · pp. 7–8
• Reported experiments achieved competitive or encouraging results against state-of-the-art, reinforcement-learning, and human-designed configurations on datasets including CIFAR-10 and CIFAR-100.Source: C. Deep Learning Architecture:… · p. 9

• Cartesian GP with a (1+λ) ES optimised symmetric convolutional autoencoders for image restoration and achieved competitive results without adversarial training.Source: D. Deep Learning Architecture:… · p. 9
• So et al. evolved transformer networks for sequence-to-sequence language tasks using a NASNet-inspired search space seeded with a known transformer model.Source: D. Deep Learning Architecture:… · p. 9
• EvoAE searches autoencoder network weights and structure simultaneously while decomposing large datasets into smaller batches to improve performance.Source: D. Deep Learning Architecture:… · p. 10
• The discriminant autoencoder is a semi-supervised autoencoder proposed for fault diagnosis.Source: D. Deep Learning Architecture:… · p. 10
• DEVDAN is an open-structure deep evolving denoising autoencoder designed for data-stream analytics and noise-perturbed input reconstruction.Source: D. Deep Learning Architecture:… · p. 10

Deep Belief Networks

• Evolutionary methods were applied to deep belief networks for image feature classification and aircraft prognostic health management.EFACV uses DBN-extracted image features, while MODBNE evolves multiple DBNs simultaneously through a decomposition-based multiobjective evolutionary algorithm.Source: E. Deep Learning Architecture: Deep… · p. 10
• A directed-acyclic-graph representation was used with evolutionary and genetic algorithms to optimize DNN structure and parameters for phoneme recognition and spoken-digit detection.The experiments used 62 GPGPUs, with restricted Boltzmann machines in training.Source: F. Other networks: LSTM, RRN, RBM · p. 10
• Differential evolution was used to identify LSTM hyperparameters for forecasting over long time spans and nonlinear continuous functions.The study describes this as the first use of differential evolution for LSTM hyperparameter selection in forecasting applications.Source: F. Other networks: LSTM, RRN, RBM · pp. 10–11
• The Semantic Learning Machine uses geometric semantic mutation in a 1 + λ hill-climbing strategy to generate child neural networks from a current best network.This representation focuses search on the current best network without imposing penalties for doing so.Source: F. Other networks: LSTM, RRN, RBM · p. 11

Final Comments

• Evolution-based methods have produced state-of-the-art DNN designs using genetic algorithms, genetic programming, evolution strategies, hybrids, and diverse representations.Reported studies range from experiments using hundreds of computers to experiments using only a few GPUs.Source: G. Final Comments · p. 12
• Most neuroevolution studies focus on deep CNN design, while autoencoders, restricted Boltzmann machines, recurrent networks, LSTMs, and deep belief machines receive less attention.The survey describes only a few works involving these other network types.Source: G. Final Comments · p. 12
• The survey tables summarize evolutionary representations, individual encodings, genetic operators, parameters, datasets, and computational effort for selected neuroevolution studies.The selected papers were chosen ad hoc to demonstrate neuroevolution in DNN architecture design and training.Source: G. Final Comments · p. 12
• Evolutionary solutions are generally considered near optimal because evolutionary algorithms provide no guarantees of convergence.Population-based evolutionary methods approximate gradients from population objectives, whereas SGD computes an exact gradient.Source: B. The critique · p. 12, A. Motivation · p. 12

Convolutional Neural Networks

• A fixed-length genetic algorithm evolved convolutional DNN weights with more than four million free parameters and achieved competitive results against several state-of-the-art methods.The comparison included DQN, policy-gradient methods, and evolution strategies.Source: C. Deep Learning Architecture:… · pp. 12–13
• Genetic algorithms were used to evolve autoencoder weights, with individuals evaluated by training-sample root mean squared error and partly updated through backpropagation.The experiments used 10 individuals, 50% elitism, and the CIFAR-10 dataset.Source: D. Deep Learning Architecture:… · p. 13
• Differentiable pattern-producing networks combine gradient-based learning with evolutionary optimization and reported superior results on MNIST.The approach uses a generic evolutionary algorithm for optimization.Source: D. Deep Learning Architecture:… · p. 13
• Evolutionary methods can tolerate sparse rewards and long time horizons, although they can struggle when optimization becomes difficult.This contrast is discussed in relation to the sensitivity and brittle convergence properties of deep reinforcement learning methods.Source: E. Other Relevant Works · p. 13
• Neuroevolution was also used to train a modular memory unit that separates memory from central computation without costly memory-management strategies.The modular memory unit is a memory-augmented recurrent architecture.Source: E. Other Relevant Works · p. 13

Final Comments

• The research community largely shifted toward stochastic gradient descent and backpropagation, but evolutionary computation remains a distinct search paradigm for DNN learning.Comparisons require caution because a neuroevolutionary generation is not directly comparable to a gradient-descent epoch.Source: F. Final Comments · pp. 13–14
• The selected DNN-training studies consistently use mutation and selection, sometimes omit crossover, and cover more varied, domain-specific datasets and problems.The survey summarizes these studies in a table of representations, operators, parameters, datasets, and computational effort.Source: Table II · p. 14, F. Final Comments · p. 14
• Evolutionary algorithms have successfully designed DNN architectures, but training those methods commonly requires days to weeks to obtain competitive results.This training-time limitation is identified across state-of-the-art architecture-design methods.Source: A. Surrogate-assisted EAs in DNNs · p. 14
• Surrogate-assisted evolutionary computation reduces fitness evaluations by estimating candidate fitness with efficient meta-models.The approach is intended to accelerate evolutionary search while maintaining accurate fitness estimates.Source: A. Surrogate-assisted EAs in DNNs · p. 14
• An ensemble-based surrogate reduced CNN training time from 33 GPU days to 10 GPU days while retaining competitive accuracy against state-of-the-art algorithms.The reported limitation is that the number of training runs needed for good prediction performance remains unknown.Source: A. Surrogate-assisted EAs in DNNs · p. 14

B. Mutations and the neutral theory

• Researchers have used selection and mutation, and often crossover, to evolve DNN architectures or train DNNs.The paper notes that crossover can be difficult depending on the encoding.Source: B. Mutations and the neutral theory · p. 14, D. Fitness Landscape Analysis of DNNs… · p. 15
• Neutrality cannot be judged as universally beneficial or harmful because its effect depends on the problem class, representation, and genetic operators.The paper reports no known neuroevolution studies focused specifically on neutrality in DNNs.Source: B. Mutations and the neutral theory · p. 15
• If neutrality is beneficial in a particular setting, it may reduce training time by avoiding evaluation of some candidate solutions.This proposed benefit is conditional on the problem, representation, and operators used.Source: B. Mutations and the neutral theory · p. 15

C. Multi-objective Optimisation

• Most reviewed neuroevolution studies optimise a single objective, although DNN configuration can require simultaneous optimisation of conflicting objectives.Multi-objective optimisation seeks trade-off solutions when no single global optimum exists.Source: C. Multi-objective Optimisation · p. 15

D. Fitness Landscape Analysis of DNNs and Well-posed Genetic Operators

• No reviewed neuroevolution study explains why a particular genetic operator is well suited to its specific DNN problem.The paper notes that standard and specialised crossover variants have both been used.Source: D. Fitness Landscape Analysis of DNNs… · p. 15

E. Standardised Scientific Neuroevolution Studies in DNNs

• Researchers use multiple DNN architectures and architectural elements, while EAs can flexibly combine methods and variants.The passage notes that this flexibility also creates many possible configuration choices.Source: E. Standardised Scientific… · p. 15
• Large datasets and stronger computational resources have enabled DNNs to address difficult problems in domains including image classification and speech processing.Source: F. Diversifying the use of benchmark… · p. 15

F. Diversifying the use of benchmark problems and DNNs

• Image classification is a primary DNN benchmark, but expanding to other domains is increasingly important for evaluating neuroevolution more broadly.The paper describes benchmark datasets as tools for comparing experimental setups across research groups.Source: F. Diversifying the use of benchmark… · p. 16
• Open benchmark datasets are critical because differing data and classification settings make published results difficult to compare.Proprietary data can further prevent direct comparison between studies.Source: F. Diversifying the use of benchmark… · p. 16

VI. CONCLUSIONS

• The survey concludes that EAs are a competitive and successful means of automatically creating and configuring DNNs.It targets researchers and practitioners in both Evolutionary Computation and Deep Learning.Source: §VI · p. 16