pso.js aims to be a clean and tiny implementation of Particle Swarm Optimization. I am currently using it in my neural network library to optimize weights. However, it's flexible enough to be used in many different kinds of optimization problems.
Check out my blog post on CodePen to see two examples on how this library can be used.
Creating a swarm is simple. All you need to do is specify the number of particles and the number of parameters you want to optimize. The last argument is the options hash.
var numParticles = 30;
var numParams = 20;
var swarm = new PSO.swarm(numParticles, numParams, {
min: -1,
max: 1,
velocityMultiplier: .1,
springCoefficient: .25,
bounceCoefficient: .25,
enableJitter: false,
jitterRatio: .01,
fitnessCompare: function(a, b) {
return a < b;
}
});The final step is to create an iteration loop.
while (true) {
for (var i in swarm.particles) {
var fitness = yourMethodOfDeterminingFitness();
swarm.particles[i].setFitness(fitness);
}
swarm.update();
}numParticles Usually the recommended number is 30. I've gotten away with using fewer, so definitely play around with it to see what works for your particular problem.
numParams The number of parameters you want to optimize. For a neural network, it would be the total number of weights and biases.
min and max Use these to define the range you want to search over. Warning: I am using a soft bounce at the edges so values may go over slightly.
velocityMultiplier How fast you want the particles to go in proportion to the search range.
springCoefficient Unless you want to slow things down so you can see the behavior more clearly, I would recommend keeping it as is. Going over will keep it from converging and going lower doesn't improve the search much.
bounceCoefficient Increase to 1 for a harder boundary. Set it to 0 if you want no boundary at all.
enableJitter and jitterRatio For more difficult problems where you find yourself getting stuck at local optima, jittering really helps. When I was using PSO for neural network regression, jittering often helped me converge with errors an order of magnitude lower than with it turned off.
fitnessCompare Just define a function that returns true when a is considered more fit than b. The default returns true when a < b, which is useful for error minimization problems.