DERIVING DATA

Swarm Design Patterns

If you read through a handful of swarm intelligence algorithms, you’ll soon see some design patterns emerging. I think it is helpful to articulate these explictly, and wield them as you study more algorithms.

Network communication topology

Interswarm communication

Intraswarm communication

Sometimes we want to apply multiple swarms to different parts of our problem. This shows up in multiobjective algorithms, as well as single objective problems conceived of as a collection of parts. If the parts are related, the performance of one swarm many influence the decisions of others, and thus definingintraswarm communication starts to emerge as a topic of interst.

Inertia Clamping

Mutation

Add noise.

Velocity Clamping

Sometimes when I implement velocity clamping in particle swarm optimization, I find myself wanting to define the functions as the “curb your enthusiasm” suite. A particle sees a major possible gain in a particular direction, and decides to throw itself whole-heartedly towards it. This is exploitation in its truest form. And yet…if we didn’t curb such enthusiasm, we’d run the risk of blowing past solutions that are even better, thereby diminishing our exploration. This wasn’t lost on Kennedy and Eberhart, as they followed up their original PSO paper with a velocity clamping solution that has since proliferated into many variets of clamping (a selection of which I’ll outline here).