Basic Machine Learning Setup in C# Console App

This document provides a detailed explanation of a single-neuron linear regression model implemented from scratch in a C# Console Application.

Model Overview

Category	Detail
Model Type	Linear Regression
Learning Algorithm	Stochastic Gradient Descent (SGD)
Architecture	Single Neuron (Perceptron without an Activation Function)
Function	Linear mapping: $\hat{y} = W_1 x_1 + W_2 x_2 + W_3 x_3 + B$
Goal	To learn the weights ($W$) and bias ($B$) that minimize the prediction error on the training data.

What the Code Accomplishes

The code demonstrates the fundamental machine learning process:

1. Training: Trains a linear model to fit a set of input-output pairs by learning the optimal weights and bias.
2. Optimization: Uses Gradient Descent to iteratively adjust these parameters to reduce the prediction error (specifically, the Mean Squared Error).
3. Prediction: After training, the model uses the learned parameters to make predictions on new, previously unseen input values.

Code Explanation: Step-by-Step Breakdown

1. Data Setup

Element	Detail	Purpose
inputs	2D array of 9 samples, 3 features each (e.g., ${1.0, 2.0, 1.0}$).	Provides the independent variables used to train the model.
outputs	1D array of 9 target values.	Provides the dependent (true) values the model must learn to predict.

2. Model Initialization

• weights (3 values): Initialized randomly. Each weight corresponds to one input feature ($x_1, x_2, x_3$).
• bias (1 value): Initialized randomly.

Why Random? Random initialization prevents the symmetry problem, ensuring that all three weights start with unique values and can learn different contributions during training.

3. Training Loop (Stochastic Gradient Descent)

The training is controlled by two nested loops:

• Outer Loop (epochs = 1000): Repeats the learning process 1000 times over the entire dataset.
• Inner Loop (Per Sample): Iterates through each of the 9 training samples.

Training Step	Formula / Concept	Explanation
Predict	$\hat{y} = W \cdot X + B$	Performs the Forward Pass: Calculates the current output based on the input and the model's current parameters.
Error	$Error = \hat{y} - y$	Calculates the residual, which is the direct measure of how wrong the current prediction is.
Weights Update	$W_{new} = W_{old} - \eta \cdot \text{Error} \cdot x_i$	Gradient Descent: Adjusts the weight by moving in the direction that decreases the loss, scaled by the learningRate ($\eta$).
Bias Update	$B_{new} = B_{old} - \eta \cdot \text{Error}$	Adjusts the model's vertical offset based on the average error.
Loss	$\text{totalLoss} += (\hat{y} - y)^2$	Accumulates the Squared Error to monitor the model's performance over the epoch.

4. Prediction

• After the 1000 epochs, the final learned weights and bias are used to predict outputs for two new input vectors: ${5, 5, 5}$ and ${20, 5, 6}$.
• The model generalizes the learned relationship (the weighted sum) to make predictions on unseen data.

5. Helper Functions

Function	Core Calculation	Role
Predict	Weighted Sum + Bias	The mathematical core of the model; computes the linear relationship.
Loss	Squared Error	The Cost Function; measures the penalty for poor predictions.