the-comfy-coder
Multiple Perceptron Neural Network
Full code can be found on my GitHub
The Project
This project aims to make a multi-perceptron neural network that recognizes the four regions of the plane to the right. It was done as a part of my CS620 AI Methods class taken in the fall of 2024.
Methods
Initializing Weights
I arbitrarily chose to have the weights initialize to a decimal between 0 and 1.0.
Calculating Output
This method takes the weights [][] and inputs[]. The length of the output array is set to the number of output neurons. Then for each output neuron i, the value of output[i] is set to 0. The weighted sum is then computed. The activation function is then applied, making the output[i] be set to 1 if the sum is more than 0 or 0 if the sum is less than 0. Once all of the output neurons have been calculated, the output array is returned.
Updating the Weights
This method modifies the weights based on the error between the target value and the output.
The formula used is w[j][i]←w[j][i]+Λ⋅e[i]⋅x[j]
Training the NN
This method trains the NN to be able to determine which region the given point is in. The regions are determined according to the table to the left, each binary pair being determined (x1,x2) such that:
- x1 is 1 if the point is above y=x and 0 if the point is below y=x
- x2 is 1 if the point is above y=-x and 0 if the point is below y=-x
The points are generated randomly within a range of -500 to 500.
Testing the NN
The operation is the same as training the NN, just without updating the weights.
Single Step Test
Allows the user to manually input a point (x,y) to see if the NN can properly determine which region it belongs in.
The Main method
Allows the user to select how many training and testing samples to use, what the learning rate should be, and then test with individual points.
Output and Findings
Testing the NN
In testing, I found that the most optimal training rate was 0.2. I arbitrarily chose 10,000 for the amount of training samples and 1,000 for the amount of testing samples.
Training Number correct: 9443
In a row: 101
Percent correct: 94.43
Testing Number correct: 998
In a row: 177
Percent correct: 99.8
Welcome to the Comfy Coder!
Recent Comments