the-comfy-coder
Kohonen NN First Attempt
In this exercise, I used a simplified Kohonen Neural Network to cluster data into sections, as seen in the example below. I completed this project as part of my CS620 AI Principals and Methods class in the fall.
The Project
The assignment was very straight forward. In Java or some other language, have the data cluster in sections as seen in the image to the right, assuming that each box is 200×200.
Methods
full code can be found on my github
Initializing Weights
The first thing you need to do in any neural network is initialize the weights of each neuron. This code creates a kohonen network of m Kohonen neurons containing n coordinates. It returns a 2d array of initialized weights. THe weights are initialized between 0 and 600 so they are all within the problem area.
Displaying the Results
This method just uses graphics to display the output, as seen later in the article.
Training Method
I chose to hard code in 500000 iterations just for simplicity, this code could easily be modified to allow for the user to choose how many iterations you’d run the program for.
While lambda is more than 0, a new point x is created and legalPoint is set to false. THis this point is still false, a new point (x,y) is made and compared to each section of the box to see if it resides inside. If the pint belongs in any of the regions, it is considered a legal point and moves on to the next step. If not the point is disposed of and a new one is randomly produced within the sample space 0-600.
If the point was legal, the closest Kohonen neuron is found and added to a list of used neurons. In our display, this changes the color from red to black. Lambda is decreased by lambda/iterations, a really small number. Every other result is displayed using the drawKohonen method previously mentioned.
Main Method
The kohonen method initializes the weights and uses the trainKohonen method to train the network.
I override the paint method to use the kohonenMethod data.
Then finally the main method, which creates an object of the kohonen class and displays the graphical output using JFrame.
Output and Findings
Using this method I was able to get the correct answer. To the right is a short video of the network in action! For this example, 500,000 iterations might be a bit overkill, as the data seems to cluster pretty quickly.
Welcome to the Comfy Coder!
Recent Comments