rylinnm/nas-with-gas

A collection of genetically-inspired algorithms that solve a nerual architecture search problem.

Neural Architecture Search using Genetic Algorithm and Evolution Strategy

Overview

This repository contains the implementation and evaluation of two algorithms: a Genetic Algorithm (GA) and an Evolution Strategy (ES), designed for Neural Architecture Search (NAS). The objective is to explore and identify high-performing neural architectures on the NAS-bench-101 benchmark. The project leverages IOHexperimenter for benchmark generation and IOHanalyzer for statistical analysis and visualization.

Algorithms

1. Genetic Algorithm (GA): Implemented in Genetic Algorithm.py, this script applies a genetic algorithm to the NAS problem. It includes functions for initialization, crossover, mutation, and various selection strategies (tournament, rank, and roulette wheel).

2. Evolution Strategy (ES): Located in Evolution Strategy.py, this script implements an evolution strategy approach for the NAS task. It provides mechanisms for parent selection, offspring generation, mutation, and survival selection.

Requirements

• Python 3.x
• nasbench
• nas_ioh
• absl
• numpy

Ensure these libraries are installed. You can install them using pip:

pip install nasbench nas_ioh absl-py numpy

Running the Code

To execute the algorithms, run the respective Python scripts from the command line:

For the Genetic Algorithm:

python Genetic Algorithm.py

For the Evolution Strategy:

python Evolution Strategy.py

Each script will perform multiple runs of the algorithm and output the best-found architecture and its performance.

Evaluation

The performance of the algorithms is evaluated based on the average best-found fitness values and AUC values over 20 independent runs, each capped at 5,000 function evaluations.