#054

A demonstration of the k-nearest neighbors algorithm implemented in javascript in a browser.

Notes

This is my working of an example from the Machine Learning with Javascript course.

Plinko is a demonstration of the k-nearest neighbors algorithm implemented in javascript in a browser.

Problem

Plinko is a simple game (like a pachinko machine). The problem to be solved: Given where ball is dropped from, can we predict which bucket it will land in?

The dependent variable (label) is the bucket. This is a classification problem as label values belong to discrete set.

The independent variables (features) that may be relevant include:

drop position
ball bounciness
ball size

How K-Nearest Neighbor Works

given a prediction point (set of features to predict label for)
collect sample data
for each observation, calculate the feature distance from prediction point
sort and take top ‘k’ records
select the most common label
that is the prediction

Multi-Dimensional KNN

When modeling with multiple features, consider:

normalize or standardise the features

Feature Normalization

normalize - 0..1
standardize - normal standard deviation around 0

Normalising with MinMax

normalized = (value - min) / (max - min)

Gauging Accuracy

Feature Selection

change in drop position - has predictable impact on bucket
change in bounciness - has impact on bucket, but not predictably

Running the Demonstration

See plinko.html or locally open plinko.html

Drop enough balls to make a sample set

optionally set bounciness and size range

asset

Click the analyze button. Results inserted at the bottom of the page: