Swarm Intelligence from social interaction
I'm hoping it's to read about Swarm Intelligence! I'm also hoping you're interested to read about the interactive dashboard side of things too so we can play with it at the end.
If that sounds like too much and you just want to play with it now, you can use it on PyViz examples here: https://particle-swarms.pyviz.demo.anaconda.com
Say you're building a house and you want to maximise the number of rooms you can fit in your plot of land, maybe saying that all rooms have to be a certain size or bigger. That's the kind of thing that optimisation algorithms are useful for.
Optimisation methods like Particle Swarm Optimisation are used when you want to find the best/optimum for some system / problem. You could just try every possible input but that might take a while so smarter people than me have invented better ways.
Let's build a dashboard in which you can control parameters of Particle Swarm Optimisation, click a target and see the little dots flock towards it. Like an interactive, 2D version of this plot on Wikipedia.
Genetic algorithm is based on genetic evolution where for each generation there is survival-of-the-fittest-style well... death. In the case of Particle Swarm Optimisation, there is the same population throughout because we want them to remember where they were when they were at their fittest. Like looking back at yourself on your wedding day or after a health kick. Each particles position is a potential solution to your problem so they're all trying to find the best position together. :heart_eyes:
In the case of Genetic Algorithm each member of the population was just a few numbers (their X and Y position), the parameters that you’re trying to optimise. In this case each particle will not just have a X and Y position, they also have a velocity. We also need a way to know how to improve the particles in our swarm...
The same as with Genetic Algorithm, we'll need to find the fittest member of the population using euclidean distance / mean squared error (which particle is closest to the target).
#collapse-hide def mean_squared_error(y_true, y_pred): return ((y_true - y_pred)**2).mean(axis=0) target_x, target_y = 0,0 def problem(soln): global target_x #using globals so we can link this to the click event later global target_y return mean_squared_error(soln, [target_x, target_y]) def assess_fitness(individual, problem): "Determines the fitness of an individual using the given problem" return problem(individual)
Each member is going to keep track of their fittest position, this can help them if they explore a worse direction, or want to tell other particles (but we'll get to that later). They also keep an ID so that we can colour them across iterations.
The big red blob is one particle which has an X and Y position, a velocity and is constantly reminiscing about it's fittest position.
Here's that in code (before we add any of the update logic).
#collapse-hide import numpy as np import pandas as pd import random from holoviews import opts, dim import holoviews as hv import panel as pn from holoviews.streams import Stream hv.extension('bokeh', logo=False)