Metaheuristic and Gradient-Based Optimization for Neural Network Training and Continuous Problems.

metANN

metANN is an R package for training feed-forward artificial neural networks using metaheuristic and gradient-based optimization algorithms. It also provides a general-purpose continuous optimization interface for benchmark functions and other numerical optimization problems.

The package supports:

• general-purpose continuous optimization,
• feed-forward multilayer perceptron training,
• regression,
• binary classification,
• multi-class classification,
• metaheuristic optimization,
• gradient-based optimization.

Main Features

metANN currently supports the following optimizers:

available_optimizers()

Currently available optimizers are:

"pso"  "de"   "ga"   "abc"  "gwo"  "woa"  "tlbo" "sboa" "sgd"  "adam"

Metaheuristic optimizers:

available_metaheuristics()

Gradient-based optimizers:

available_gradient_optimizers()

Installation

You can install the development version of metANN from GitHub using:

# install.packages("remotes")
remotes::install_github("burakdilber/metANN")

After installation, load the package:

library(metANN)

Replace YOUR_GITHUB_USERNAME with your GitHub username after uploading the package to GitHub.

General-Purpose Optimization

The met_optimize() function can be used for continuous optimization problems.

Metaheuristic optimizers require only an objective function.

Gradient-based optimizers require both an objective function and its gradient.

Example 1: Sphere Function with SBOA

sphere <- function(x) {
  sum(x^2)
}

result_sboa <- met_optimize(
  fn = sphere,
  optimizer = optimizer_sboa(
    pop_size = 30,
    max_iter = 100
  ),
  lower = rep(-10, 10),
  upper = rep(10, 10),
  seed = 123,
  verbose = FALSE
)

result_sboa
summary(result_sboa)
coef(result_sboa)

plot(result_sboa)
plot(result_sboa, log = TRUE)

Example 2: Rastrigin Function with SBOA

rastrigin <- function(x) {
  10 * length(x) + sum(x^2 - 10 * cos(2 * pi * x))
}

result_rastrigin <- met_optimize(
  fn = rastrigin,
  optimizer = optimizer_sboa(
    pop_size = 40,
    max_iter = 150
  ),
  lower = rep(-5.12, 5),
  upper = rep(5.12, 5),
  seed = 123,
  verbose = FALSE
)

result_rastrigin
summary(result_rastrigin)

plot(result_rastrigin)
plot(result_rastrigin, log = TRUE)

Gradient-Based General Optimization

Gradient-based optimizers such as SGD and Adam require a gradient function supplied via the gr argument.

Example 3: Sphere Function with Adam

sphere <- function(x) {
  sum(x^2)
}

grad_sphere <- function(x) {
  2 * x
}

result_adam <- met_optimize(
  fn = sphere,
  gr = grad_sphere,
  optimizer = optimizer_adam(
    learning_rate = 0.1,
    epochs = 100
  ),
  lower = rep(-5, 5),
  upper = rep(5, 5),
  initial = rep(4, 5),
  seed = 123,
  verbose = FALSE
)

result_adam
summary(result_adam)
coef(result_adam)

plot(result_adam)
plot(result_adam, log = TRUE)

Example 4: Rastrigin Function with Adam

rastrigin <- function(x) {
  10 * length(x) + sum(x^2 - 10 * cos(2 * pi * x))
}

grad_rastrigin <- function(x) {
  2 * x + 20 * pi * sin(2 * pi * x)
}

result_adam_rastrigin <- met_optimize(
  fn = rastrigin,
  gr = grad_rastrigin,
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 500
  ),
  lower = rep(-5.12, 5),
  upper = rep(5.12, 5),
  initial = rep(3, 5),
  seed = 123,
  verbose = FALSE
)

result_adam_rastrigin
summary(result_adam_rastrigin)

plot(result_adam_rastrigin)

Since the Rastrigin function is multimodal, gradient-based optimizers may converge to local minima depending on the initial point and learning rate.

Feed-Forward Neural Network Training

The met_mlp() function is a convenient wrapper for training feed-forward multilayer perceptrons.

The task can be selected manually using:

task = "regression"
task = "classification"

or detected automatically using:

task = "auto"

When task = "auto":

• numeric response variables are treated as regression,
• factor, character, or logical response variables are treated as classification.

Example 5: MLP Regression with SBOA

fit_reg_sboa <- met_mlp(
  formula = Petal.Width ~ Sepal.Length + Sepal.Width + Petal.Length,
  data = iris,
  hidden_layers = c(5),
  activation = "relu",
  optimizer = optimizer_sboa(
    pop_size = 20,
    max_iter = 30
  ),
  seed = 123,
  verbose = FALSE
)

fit_reg_sboa
summary(fit_reg_sboa)

evaluate(fit_reg_sboa, newdata = iris)

predict(fit_reg_sboa, iris[1:5, ], type = "response")

plot(fit_reg_sboa)

Example 6: MLP Regression with Adam

fit_reg_adam <- met_mlp(
  formula = Petal.Width ~ Sepal.Length + Sepal.Width + Petal.Length,
  data = iris,
  hidden_layers = c(5),
  activation = "relu",
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 100,
    batch_size = 32
  ),
  seed = 123,
  verbose = FALSE
)

fit_reg_adam
summary(fit_reg_adam)

evaluate(fit_reg_adam, newdata = iris)

predict(fit_reg_adam, iris[1:5, ], type = "response")

plot(fit_reg_adam)

Binary Classification

For binary classification, metANN uses a one-unit output layer with sigmoid activation by default.

If loss = NULL, the default loss is binary cross-entropy.

Example 7: Binary Classification with SBOA

iris_bin <- iris

iris_bin$IsSetosa <- factor(
  ifelse(iris_bin$Species == "setosa", "setosa", "other")
)

fit_bin_sboa <- met_mlp(
  formula = IsSetosa ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width,
  data = iris_bin,
  hidden_layers = c(5),
  activation = "relu",
  optimizer = optimizer_sboa(
    pop_size = 20,
    max_iter = 30
  ),
  seed = 123,
  verbose = FALSE
)

fit_bin_sboa
summary(fit_bin_sboa)

evaluate(fit_bin_sboa, newdata = iris_bin)

predict(fit_bin_sboa, iris_bin[1:5, ], type = "class")

predict(fit_bin_sboa, iris_bin[1:5, ], type = "prob")

plot(fit_bin_sboa)

Example 8: Binary Classification with Adam

iris_bin <- iris

iris_bin$IsSetosa <- factor(
  ifelse(iris_bin$Species == "setosa", "setosa", "other")
)

fit_bin_adam <- met_mlp(
  formula = IsSetosa ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width,
  data = iris_bin,
  hidden_layers = c(5),
  activation = "relu",
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 100,
    batch_size = 32
  ),
  seed = 123,
  verbose = FALSE
)

fit_bin_adam
summary(fit_bin_adam)

evaluate(fit_bin_adam, newdata = iris_bin)

predict(fit_bin_adam, iris_bin[1:5, ], type = "class")

predict(fit_bin_adam, iris_bin[1:5, ], type = "prob")

plot(fit_bin_adam)

For binary classification, type = "prob" returns a two-column probability matrix.

Multi-Class Classification

For multi-class classification, metANN uses a softmax output layer by default.

If loss = NULL, the default loss is categorical cross-entropy.

Example 9: Multi-Class Classification with SBOA

fit_multi_sboa <- met_mlp(
  formula = Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width,
  data = iris,
  hidden_layers = c(6),
  activation = "relu",
  optimizer = optimizer_sboa(
    pop_size = 25,
    max_iter = 40
  ),
  seed = 123,
  verbose = FALSE
)

fit_multi_sboa
summary(fit_multi_sboa)

evaluate(fit_multi_sboa, newdata = iris)

predict(fit_multi_sboa, iris[1:5, ], type = "class")

predict(fit_multi_sboa, iris[1:5, ], type = "prob")

plot(fit_multi_sboa)

Example 10: Multi-Class Classification with Adam

fit_multi_adam <- met_mlp(
  formula = Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width,
  data = iris,
  hidden_layers = c(6),
  activation = "relu",
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 200,
    batch_size = 32
  ),
  seed = 123,
  verbose = FALSE
)

fit_multi_adam
summary(fit_multi_adam)

evaluate(fit_multi_adam, newdata = iris)

predict(fit_multi_adam, iris[1:5, ], type = "class")

predict(fit_multi_adam, iris[1:5, ], type = "prob")

plot(fit_multi_adam)

For multi-class classification, type = "prob" returns a probability matrix with one column per class.

Prediction Types

For regression models:

predict(fit_reg_adam, newdata = iris[1:5, ], type = "response")

For classification models:

predict(fit_bin_adam, newdata = iris_bin[1:5, ], type = "class")

predict(fit_bin_adam, newdata = iris_bin[1:5, ], type = "prob")

predict(fit_bin_adam, newdata = iris_bin[1:5, ], type = "response")

For classification models:

• type = "class" returns predicted class labels,
• type = "prob" returns predicted class probabilities,
• type = "response" returns class labels by default.

Network Visualization

The plot_network() function visualizes the architecture of a fitted metANN model or an MLP architecture object.

plot_network(fit_multi_adam)

It displays input, hidden, and output layers, including the number of neurons and activation functions.

You can also plot an architecture object directly:

arch <- mlp_architecture(
  input_dim = 4,
  layers = list(
    dense_layer(6, activation = "relu"),
    dense_layer(3, activation = "softmax")
  )
)

plot_network(arch)

Model Evaluation

The evaluate() function computes performance metrics on new data.

For regression models:

evaluate(fit_reg_adam, newdata = iris)

evaluate(
  fit_reg_adam,
  newdata = iris,
  metrics = c("rmse", "mae", "r2")
)

For binary classification models:

evaluate(fit_bin_adam, newdata = iris_bin)

evaluate(
  fit_bin_adam,
  newdata = iris_bin,
  metrics = c("accuracy", "precision", "recall", "f1")
)

For multi-class classification models:

evaluate(fit_multi_adam, newdata = iris)

evaluate(
  fit_multi_adam,
  newdata = iris,
  metrics = c("accuracy", "precision", "recall", "f1")
)

x-y Interface

In addition to the formula-data interface, metANN also supports an x-y interface.

Example 11: Regression with x-y Interface

x_reg <- as.matrix(
  iris[, c("Sepal.Length", "Sepal.Width", "Petal.Length")]
)

y_reg <- iris$Petal.Width

fit_xy_reg <- met_mlp(
  x = x_reg,
  y = y_reg,
  hidden_layers = c(5),
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 100,
    batch_size = 32
  ),
  seed = 123,
  verbose = FALSE
)

fit_xy_reg
summary(fit_xy_reg)

evaluate(
  fit_xy_reg,
  newdata = x_reg,
  y_true = y_reg
)

Example 12: Binary Classification with x-y Interface

iris_bin <- iris

iris_bin$IsSetosa <- factor(
  ifelse(iris_bin$Species == "setosa", "setosa", "other")
)

x_bin <- as.matrix(
  iris_bin[, c("Sepal.Length", "Sepal.Width", "Petal.Length", "Petal.Width")]
)

y_bin <- iris_bin$IsSetosa

fit_xy_bin <- met_mlp(
  x = x_bin,
  y = y_bin,
  hidden_layers = c(5),
  optimizer = optimizer_adam(
    learning_rate = 0.01,
    epochs = 100,
    batch_size = 32
  ),
  seed = 123,
  verbose = FALSE
)

fit_xy_bin
summary(fit_xy_bin)

evaluate(
  fit_xy_bin,
  newdata = x_bin,
  y_true = y_bin
)

Optimizer Information

The optimizer_info() function provides a short summary of an optimizer.

optimizer_info("sboa")

optimizer_info(
  optimizer_adam(
    learning_rate = 0.01,
    epochs = 200,
    batch_size = 32
  )
)

Example output includes:

• optimizer name,
• full optimizer name,
• optimizer type,
• whether it requires gradients,
• supported interfaces,
• main parameters,
• current parameter values.

Available Activation Functions

available_activations()

Currently available activation functions are:

"linear" "sigmoid" "tanh" "relu" "leaky_relu" "softmax"

Available Loss Functions

available_losses()

Currently available loss functions are:

"mse" "mae" "huber" "log_cosh" "binary_crossentropy" "crossentropy"

Available Metrics

available_metrics()

Currently available metrics are:

"mse" "rmse" "mae" "r2" "accuracy" "precision" "recall" "f1"

Supported Tasks

The current development version supports:

Notes

In the current development version:

• metaheuristic optimizers can train MLPs without requiring gradients,
• SGD and Adam can optimize differentiable benchmark functions when a gradient function is supplied,
• SGD and Adam can train MLPs for regression, binary classification, and multi-class classification,
• binary classification uses sigmoid output and binary cross-entropy by default,
• multi-class classification uses softmax output and cross-entropy by default,
• classification metrics are computed using macro-averaged precision, recall, and F1.

References

The optimization algorithms and neural network components implemented in metANN are based on the following key references:

• Kennedy, J., and Eberhart, R. (1995). Particle Swarm Optimization. Proceedings of ICNN’95 - International Conference on Neural Networks, 4, 1942–1948.
• Storn, R., and Price, K. (1997). Differential Evolution – A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces. Journal of Global Optimization, 11, 341–359.
• Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley.
• Karaboga, D., and Basturk, B. (2007). A Powerful and Efficient Algorithm for Numerical Function Optimization: Artificial Bee Colony (ABC) Algorithm. Journal of Global Optimization, 39, 459–471.
• Mirjalili, S., Mirjalili, S. M., and Lewis, A. (2014). Grey Wolf Optimizer. Advances in Engineering Software, 69, 46–61.
• Mirjalili, S., and Lewis, A. (2016). The Whale Optimization Algorithm. Advances in Engineering Software, 95, 51–67.
• Rao, R. V., Savsani, V. J., and Vakharia, D. P. (2011). Teaching-Learning-Based Optimization: A Novel Method for Constrained Mechanical Design Optimization Problems. Computer-Aided Design, 43, 303–315.
• Fu, Y., Liu, D., Chen, J., and He, L. (2024). Secretary Bird Optimization Algorithm: A New Metaheuristic for Solving Global Optimization Problems. Artificial Intelligence Review, 57, 123.
• Kingma, D. P., and Ba, J. (2015). Adam: A Method for Stochastic Optimization. International Conference on Learning Representations.
• Nair, V., and Hinton, G. E. (2010). Rectified Linear Units Improve Restricted Boltzmann Machines. Proceedings of the 27th International Conference on Machine Learning, 807–814.
• Bridle, J. S. (1990). Probabilistic Interpretation of Feedforward Classification Network Outputs, with Relationships to Statistical Pattern Recognition. In Neurocomputing: Algorithms, Architectures and Applications, 227–236.

Metaheuristic-Based Neural Network Training

The neural network training functionality in metANN is related to the literature on metaheuristic-based training of feed-forward neural networks. Representative studies include:

• Montana, D. J., and Davis, L. (1989). Training Feedforward Neural Networks Using Genetic Algorithms. Proceedings of the 11th International Joint Conference on Artificial Intelligence, 762–767.

• Ilonen, J., Kamarainen, J.-K., and Lampinen, J. (2003). Differential Evolution Training Algorithm for Feed-Forward Neural Networks. Neural Processing Letters, 17, 93–105. doi:10.1023/A:1022995128597

• Karaboga, D., and Ozturk, C. (2009). Neural Networks Training by Artificial Bee Colony Algorithm on Pattern Classification. Neural Network World, 19(3), 279–292.

• Mirjalili, S. (2015). How Effective is the Grey Wolf Optimizer in Training Multi-Layer Perceptrons. Applied Intelligence, 43, 150–161. doi:10.1007/s10489-014-0645-7

• Dilber, B., and Özdemir, A. F. (2026). A novel approach to training feed-forward multi-layer perceptrons with recently proposed secretary bird optimization algorithm. Neural Computing and Applications, 38(5). doi:10.1007/s00521-026-11874-x

Citation

If you use metANN in academic work, please cite the package and the related optimization algorithms where appropriate.

Authors

metANN is developed and maintained by:

• Burak Dilber
• A. Fırat Özdemir

License

This package is released under the MIT license.