For training a neural-network, we have an awesome back-propagation algorithm that allows us automatically tweak our weights and biases to fit our dataset. This is just a project came out of curiosity -- to test for another method that is possible to tweak the network without depending on any model-based algorithm.
To be honest, back-propagation is still the winning choice here. This is just a proof-of-concept project, which has proved that randomness of genetic algorithm is still possible to let the network learn, albeit very very slow learning.
Please be noted, for large dimension of data (eq: mnist/cifar-10), back-progapation wins the competition by tenfolds. That is why for this project I chose to use Iris dataset, as it is small enough for me to conduct an experiment.
- Install dependencies
pip install -U scikit-learn numpy pandas
- Run the project
python neural-net-ga.py
