shawn-bo / pso-stuff

Various PSO implementations. Check https://arxiv.org/abs/2006.09875

EMPSO, FCPSO and CRNGs

Contains implementation of EMPSO by Mohapatra et. al.

Also contains implementation of three variants of FCPSO -

• FCPSO is SMPSO adapted to single-objective problems.
• FCPSO-em-Beta is SMPSO with exponentially-averaged momentum drawn from a uniform range of the momentum parameter that makes it fairly constricted.
• FCPSO-em-Omega is also a fairly constricted PSO algorithm, but the momentum parameter distribution is non-uniform

There is also an implementation of Chaotic Random Number Generators (CRNG) in the module chaosGen.py. Below is an example of the CRNG.

The aim of EMPSO is for gradient approximation in neural networks. We have tested EMPSO on simple 1 dimensional objective functions for the accuracy of its gradient. It has worked well on a highly non-linear and multimodal Rastrigin function. Please check this notebook for the demo.

This repository also contains the implementation of a novel PSO algorithm named Reverse-Informed Locally-Searched PSO. A demo of this PSO variant can be found in this notebook. The special feature of this variant is that it is capable of detecting multiple equally fit global optima. Please refer to the details of which can be found here.