Drawing With De-Optimized Genetic Algorithms

By Brian Maresso

Background

Genetic Algorithms are remarkable for their ability to optimize complex solutions while being relatively simple in their implementation.  So long as a problem can evaluate solutions for correctness by some arbitrary metric, genetic algorithms can (theoretically) find the optimal solution- even when the solution is far from what a human would expect (The 2007 NASA ST5 Spacecraft Evolved Antenna is a great example).

I’ve been using genetic algorithms for years, primarily as a training mechanism for neural networks (a technique known as neuroevolution).  In these applications, we’re typically looking for precise solutions to minimize error.  Instead, this project will focus on de-optimizing genetic algorithms to intentionally introduce error into otherwise ‘perfect’ solutions.  Of course, this is hardly practical for most computer science applications.  Instead, this process of de-optimization will be used as an image processing technique for various abstract effects.

• Part 1: Coming Soon!