Evolving neural network code based on existing human code.

Our implementation requires the following.

• Python 2.7
• PyTorch 0.4.0

Errors have been reported when trying to run the code on Python 3 or on higher or lower versions of PyTorch. It also assumes there is a GPU in the system, which the author guesses is a dependency.

The genetic algorithm is implemented in the GACode directory.

One must make the directories specified in code_path and img_path in the ga.py code. These are directories where the resulting architectures and Pareto front images will be saved, respectively.

All the following regards the ga.py code. The baseline models to place in the initial population can be modified by changing the components in the INIT_POP variable. The amount of growth time can be specified using the growth_time variable. The population size can be specified using the pop_size variable. The number of generations can be specified by setting the number in the breed function to the desired number.

After all parameters have been specified, execution is simple:

python ga.py

The best-performing networks' names and their measured performance is saved in the file name specified in the result_file variable. The code from each generation is saved in the directory specified in the code_path variable. A plot of best-performing networks can be collected from the img_path.

This project was done as part of a term project in the AI Based Software Engineering course at KAIST in Autumn 2018.