Genetic Algorithm for Network Topology

Usage

usage: main.py [-h] [--n N] [--e E] [--p P] [--a A] [--c C] [--w W]

A genetic algorithm that generates network topologies based on user input

optional arguments:
  -h, --help  show this help message and exit
  --n N       The number of nodes in the network
  --e E       The number of edges in the network
  --p P       Parent pool size for the GA
  --a A       Alpha value for evaluation: [0, 1]
  --c C       Whether the graph should be connected (default: true)
  --w W       Whether the graph should be weighted

Pipeline

This GA will be split into the following components, with each interaction described:

main.py

Purpose: To initiate and to end the entire process
Input: User execution or kill notification, any custom parameters (such as alpha)
Output: Best networks and their scores

generator.py

Purpose: Generate an initial generation of seed networks
Input: Custom parameters for network constraints (if any)
Output: A network described in an object

iterator.py

Purpose: Begin and maintain a loop of evaluating and mutating networks
Input: Seed network generation from generator
Output: A best network given certain specified stopping mechanisms

evaluator.py

Purpose: Runs fitness simulations on networks to evaluate them
Input: A network
Output: A score

mutator.py

Purpose: Mutate the networks that performed the best
Input: A set of networks (or a network)
Output: A new generation of mutated networks

Changing Evaluation Mechanism

This GA can run on SECON metrics or on the custom metrics. Make these changes in evaluator.py. Uncomment/comment lines 52, 54 for efficiency. Or lines 79, 82-86, 89, 92 for robustness.

Seeing results

If you run the GA and see images for each generation populating the img/ folder, then you can run the following command to convert all of the images into a video:

convert -quality 100 -delay 10x100  *.png gen.mpeg

TODO

There are still a couple of things that need to be worked on. For example, a multi-starting point needs to be added as an additional way to overcome local optima. In terms of extensibility, the directedness and weightedness have not been tested since all of the research results have been done on unweighted, undirected graphs. However, the options are there and should be able to be quickly debugged should any issues arise.

jacksonchen / network_genetic