CNN_genetic_algorithm

The only problem i previously used Genetic Algorithm(GA) was in knapsack problem. So I got a really cool idea of classifying the IRIS dataset usin CNN, where GA is used to train the weights.

Training a CNN with genetic algorithm using the classic Iris dataset

Requirements:

$ pip install tensorflow
$ pip install keras
$ pip install numpy
$ pip install pandas
$ pip install scikit-learn
$ pip install pygad

This blog is used as the main inspiration in training the IRIS dataset using a Genetic Algorithm based CNN architecture.

This is the documentation to read more about PyGAD, PyGAD is a python library for implementing genetic algorithm, and it also supports keras integration.

This kaggle post provides a detailed visual representation of the IRIS dataset.

Network Architecture of the CNN used:

Layer (type)                 Output Shape              Param #   
=================================================================
input_10 (InputLayer)        [(None, 4)]               0         
_________________________________________________________________
dense_25 (Dense)             (None, 16)                80        
_________________________________________________________________
dense_26 (Dense)             (None, 8)                 136       
_________________________________________________________________
dense_27 (Dense)             (None, 3)                 27        
=================================================================
Total params: 243
Trainable params: 243
Non-trainable params: 0
_________________________________________________________________

Hyperparameters used in training:

There are two basic hyperparameters used in the GA model, those are: num_generations, and num_parents_mating. Where the parameter values used are:

num_generations = 250
num_parents_mating = 5

Absolute error

The model reaches an Absolute Error : 0.026681786

Fitness vs Generation graph: