Happy Sunday and welcome to InsideAI! I'm Rob May, Founder and CTO at Dianthus, and a venture partner at PJC. I've been writing the commentary here for 6 years to help share things I'm seeing in the AI ecosystem. This week, I want to share an older AI technology that I'm seeing a little bit more of in the market - evolutionary algorithms (sometimes referred to as genetic algorithms).
Evolutionary algorithms are modeled off evolution. You take an initial possible solution to a problem, randomly generate a bunch of variations on it, test all the variations against the problem, keep the best solution, and throw away the others. Then run another cycle with the new solution and variants of that one.
Evolutionary algorithms are powerful because they don't require a lot of data. They work differently than neural networks because they aren't "trained" the way neural nets are. The downsides are, evolutionary algorithms are often slow to run, and don't work on many classes of problems. And the biggest issue is it's often difficult to encode a problem effectively into an evolutionary algorithm model that can help.
One of the most famous success stories of evolutionary algorithms is the evolved antenna, a design that came from an algorithm that created a very strange looking antenna no human would have thought to design. But the electromagnetic properties of the antenna work well.
Evolutionary algorithms are good for projects where you need to explore a new and open problem-solution space and trial and error may be your best bet. If you've hit the wall with other methods, it might be time to give EAs a try. I believe you will see a lot more of them in coming years.
Thanks for reading.
@robmay
Comments
Post a Comment