GA-ANN: Genetic Algorithm-Optimized Neural Networks

• GA-ANN is a hybrid approach that integrates genetic algorithms with artificial neural networks to optimize hyperparameters, network topology, weights, and feature selection.
• It employs both discrete and continuous search strategies, including binary encoding and real-valued optimization, to fine-tune model performance and reduce training time.
• Applied across fields from cosmology to industrial systems, GA-ANN has demonstrated improved error metrics, enhanced accuracy, and faster convergence compared to traditional methods.

Genetic Algorithm-Optimized Artificial Neural Network (GA-ANN) denotes a family of hybrid methods in which a genetic algorithm is coupled to an artificial neural network so that evolutionary search operates on some part of the learning problem—hyperparameters, architecture, weights, input variables, feature subsets, or expensive evaluation procedures—while the neural network serves either as the predictive model or as a search accelerator. In the literature surveyed here, GA-ANN spans hyperparameter optimization for feedforward networks, CNNs, RNNs, and capsule networks; layer-wise evolution of deep autoencoder weights; wrapper-based variable and feature selection for MLPs; and surrogate-assisted evolutionary search in simulation-based design (Vose et al., 2019, Gómez-Vargas et al., 2022, David et al., 2017, Al-Batah et al., 11 Jun 2025, 0706.1051, Paape et al., 2024).

1. Conceptual scope

The optimization target varies substantially across published GA-ANN systems. In cosmology, GA is used as an external hyperparameter optimizer over number of hidden layers, number of nodes per layer, learning rate, and batch size, while ANN weights are still learned by Adam (Gómez-Vargas et al., 2022). In TinyFace recognition, GA is explicitly a wrapper feature-selection mechanism preceding MLP training rather than a direct optimizer of MLP weights or hidden-layer parameters (Al-Batah et al., 11 Jun 2025). In textile defect classification and manhole-gas estimation, GA is used to train or optimize ANN parameters without relying on differentiability of the activation or error surface (Habib et al., 2012, Ojha et al., 2012). In other hybrids, the direction is reversed: a neural network accelerates the GA by proposing chromosomes or approximating expensive objective evaluations (Acharya et al., 2024, Paape et al., 2024).

This breadth makes GA-ANN a category of coupling strategies rather than a single algorithm. Some methods use GA to search over discrete ANN design choices; some use GA for direct continuous parameter search; some wrap ANN training inside a combinatorial search over sensors or extracted features; and some use a neural model as a learned prior or surrogate within the evolutionary loop (0706.1051, Hebbar, 2023, Sheikhan et al., 2017).

The abstract of "Deep Genetic Network" states that the proposed architecture uses genetic algorithms along with deep neural networks to optimize hyperparameters in affine, convolutional and recurrent layers, and that the network learns to optimize its hyperparameters by itself rather than depending on a person to explicitly set the values. However, no PDF or source is available for (Choudhury et al., 2018)v2, so the proposed architecture, fitness function, datasets, baselines, and results cannot be recovered beyond the abstract (Choudhury et al., 2018).

2. Optimization objects and representation schemes

Chromosome design follows the object being optimized. For discrete hyperparameter search, the cosmology study uses binary encoding of hyperparameter choices, with each individual corresponding to one combination of layers, nodes, learning rate, and batch size (Gómez-Vargas et al., 2022). The GA aided by a bi-generative adversarial network represents convolutional layers as

$$prm_i^{conv^j}=[1/0\text{ (conv. layer exists or not)}, \text{kernel size}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for max pooling)}]$$

and dense layers as

$$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$

while the Bi-GAN proposes the number of filters and neurons from a larger range than a finite preset candidate set (Lu et al., 2018).

When the objective is direct ANN parameter optimization, chromosomes encode weight sets or flattened parameter vectors. In the deep autoencoder method, each chromosome corresponds to one complete set of weights for the current autoencoder layer and fitness is defined as

$f_i = \frac{1}{RMSE_i}.$

Because training is greedy and layer-wise, the chromosome does not encode the entire deep network at once (David et al., 2017). In the single-hidden-layer feedforward setting, the trainable parameters are collected into

$$\mathbf{w} = \left(w_1, w_2, \ldots, w_m\right)^T \in \mathbb{R}^m,$$

and the optimization target is the empirical loss

$$\text{MSE}(\mathbf{w}) = \frac{1}{N} \sum_{i=1}^{N} \left( y_i - \hat{y}_i(\mathbf{w}) \right)^2.$$

The architecture is fixed for a given experiment; GA acts as a real-valued optimizer over weights and biases rather than as a topology-evolving method (Kashyap et al., 17 Jun 2025).

Other systems retain a neural model in the loop but shift the genotype away from full-network encoding. The manhole-gas estimator treats each synaptic weight as a gene in IEEE 754 32-bit floating-point format, with chromosome length $32N$ bits for $N$ weights, and minimizes

$SSE = \frac{1}{2}\sum \sum (O_{pi} - t_{pi})^2$

over all patterns and output nodes (Ojha et al., 2012). Wrapper-based approaches instead encode subsets of variables or extracted features and use ANN performance on those subsets as fitness, so the chromosome describes the ANN input space rather than its full parameter state (0706.1051, Al-Batah et al., 11 Jun 2025).

3. Evolutionary dynamics and hybrid learning loops

GA-ANN seldom appears as a pure replacement for conventional learning; hybridization is the recurrent pattern. A prominent example is tri-parent population-based training, where two parents contribute hyperparameters and a third contributes model parameters. The offspring is formed by

$$c.hparams \leftarrow mutate(crossover(a.hparams, b.hparams)), \qquad c.params \leftarrow \alpha.params,$$

thereby decoupling weights and biases from hyperparameters during sexual reproduction and interleaving SGD with GA across epochs (Vose et al., 2019).

Locality-controlled evolutionary dynamics constitute another line of development. Cellular genetic algorithms place individuals on a toroidal 2D grid, restrict interaction to local neighborhoods, and support synchronous or asynchronous updates with elitist replacement. The stated purpose is to control the spread of solutions through the population and preserve genetic diversity for longer time (Silva et al., 2021).

Search guidance can also be layered on top of the evolutionary process. The MCTS-guided method treats GA evolution as a tree, uses UCT with

$UCT = Q + UCB1,$

defines actions as genetic operations of selection, crossover, and mutation, and performs evolutionary rollouts for 10 generations to estimate downstream utility of a candidate (Hebbar, 2023). The air-mobility framework trains a simple NN with two hidden layers to map

$$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$0

then injects NN-generated chromosomes into the GA population; random addition of NN-generated parents is reported to perform better than replacement of the least-fit individuals (Acharya et al., 2024).

Some controllers distribute optimization across multiple metaheuristics. In adaptive congestion control, GA selects the number of hidden neurons in an RBF controller, while PSO tunes the output weights and the integral gain of the improved RBF variant to minimize

$$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$1

This is a GA-ANN system in which structural search and continuous parameter search are explicitly split (Sheikhan et al., 2017).

4. Neural architectures and application domains

The neural architectures used in GA-ANN research are highly heterogeneous. Feedforward MLPs dominate the cosmology case studies, where GA replaces exhaustive grid search for model-independent reconstruction of the distance modulus from the JLA compilation of 740 Type Ia supernovae, inference of the equation of state for a quintessence model on a dataset of 2800 combinations, and photometric redshift prediction from six photometric bands $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$2 (Gómez-Vargas et al., 2022). CNNs and related deep architectures appear in ModelNet40 optimization, MNIST autoencoding, and TinyFace recognition after VGG19 feature extraction (Lu et al., 2018, David et al., 2017, Al-Batah et al., 11 Jun 2025).

Control and systems applications often use domain-specific ANN forms rather than generic MLPs. Active queue management employs Gaussian RBF and improved RBF neural controllers in a nonlinear delayed TCP model, with controller output interpreted as a packet drop or mark probability bounded in $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$3 (Sheikhan et al., 2017). Fractional-order dynamics uses GA to fit an equivalent second-order model and then trains a multilayer feed-forward ANN to predict optimal pseudo-damping and meta-damping from the fractional-order descriptor, with the best reported ANN structure being a single hidden layer with 5 neurons and logsig activation (Das et al., 2012).

Industrial and operational optimization problems frequently use GA-ANN in wrapper or surrogate form. In a liquid fed ceramic melter, GA searches the $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$4 nonempty subsets of 20 thermocouple sensors to determine ANN inputs for predicting melt chamber level (0706.1051). In production-system topology optimization, GA searches a binary categorical encoding of component presence and port-to-port connections while a feedforward neural network acts as a surrogate for expensive discrete-event simulation (Paape et al., 2024). In pre-disaster evacuation planning, a neural-network-accelerated GA selects destination airports with 10-bit chromosomes for hourly evacuation scheduling using FAA TFMSC and FAA ASPM data (Acharya et al., 2024).

Other studies remain close to direct neuroevolution. GA-assisted deep autoencoders train a 5-layer structure with layer sizes 784–500–250–100–50 in greedy layer-wise fashion on MNIST (David et al., 2017). GRWC combines GA with Random Weight Change learning, tests the method on MNIST, and then removes all weak weights below the threshold value 0.1, reducing the total weights by almost half for hardware-oriented implementation (Sarker et al., 2019).

5. Reported empirical record

Reported outcomes differ by task, metric, and optimization target, but several studies provide direct numerical comparisons.

Setting	Reported comparison	Source
ModelNet40, 3D CNN	ShapeNet 0.8417; Small-set GA 0.8294; Large-set GA 0.3641; GA + Bi-GAN 0.8520	(Lu et al., 2018)
TinyFace with VGG19 features and MLP	Baseline MLP 0.478; GA feature selection + MLP 0.545; PCA + MLP 0.125	(Al-Batah et al., 11 Jun 2025)
Deep autoencoder on MNIST features	Traditional autoencoder 1.85% classification error; GA-assisted 1.44%	(David et al., 2017)
Cosmology: JLA, quintessence, photo-$$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$5	GA improves over grid in all three cases; e.g. quintessence MSE $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$6 to $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$7, MAE 0.0174 to 0.0020, $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$8 0.9873 to 0.9998	(Gómez-Vargas et al., 2022)
Population-based training	CANDLE P3B1 time 118.6 s to 25.8 s; LeNet 164.5 s to 32.4 s; ResNet-20 error about 8% to just over 7%	(Vose et al., 2019)
Production topology optimization	Mean evaluations before optimum: unassisted GA 576; feedforward ANN-assisted GA 265	(Paape et al., 2024)

Additional task-specific results broaden this empirical picture. In textile defect classification, the selected 4-26-6 network reaches 91.75% accuracy with fitness $$prm_i^{dense^k}=[1/0\text{ (dense layer exists or not)}, \text{activation func.}, 1/0\text{ (for batch norm.)}, 1/0\text{ (for dropout)}],$$9 after GA tuning of population size, crossover rate, mutation rate, and hidden-layer size (Habib et al., 2012). In the gas-mixture estimator, performance improves with population size and generations; with 100 chromosomes and 1000 iterations, the error drops below 3% (Ojha et al., 2012).

These results suggest that GA-ANN can yield gains in several distinct regimes: hyperparameter search over discrete design spaces, sparse or more discriminative representation learning, time-to-accuracy reduction in population-based training, and reduction in expensive ground-truth evaluations through neural surrogates.

6. Limitations, misconceptions, and methodological boundaries

GA-ANN results are not uniformly dominant. The single-hidden-layer feedforward study finds that GA with crossover consistently beats GA without crossover and is much better than random search, but hybrid PSO-SGD is generally strongest on small and medium networks, and the advantage of GA diminishes as network dimensionality grows (Kashyap et al., 17 Jun 2025). The cosmology study likewise notes that GA remains stochastic and results depend on random initialization and pseudo-random training behavior (Gómez-Vargas et al., 2022).

Several papers report domain-specific failure modes or practical costs. DIGA may require up to 20,000 iterations, converges more slowly than gradient descent, and on the four-layer cat/non-cat setting gradient descent reaches 80% test accuracy versus 76% for DIGA despite DIGA’s lower training cost, which the paper interprets as overfitting (Truong et al., 24 Apr 2025). The cellular GA study identifies high runtime and processing cost due to nested loops, Matlab implementation, and the likely need for parallel programming in a more suitable implementation (Silva et al., 2021).

A common simplification is to equate GA-ANN with direct evolution of network weights. The literature is broader: GA can be a wrapper feature selector for an MLP, a variable selector for ANN-based industrial modeling, or one component of a surrogate-assisted optimizer where a neural network predicts simulation-based fitness (Al-Batah et al., 11 Jun 2025, 0706.1051, Paape et al., 2024). This suggests that GA-ANN is better understood as a reusable coupling pattern between evolutionary search and neural modeling than as a single canonical algorithm.

Methodological completeness also varies across the record. The abstract of "Deep Genetic Network" presents a GA-based deep architecture for self-optimizing hyperparameters and states that it works well in affine, convolutional and recurrent layers, but the unavailable PDF/source means that methods, equations, datasets, baselines, and results cannot presently be verified beyond the abstract (Choudhury et al., 2018).