🏎️ NeuroDrive

NeuroDrive is an automated neural network benchmarking suite designed to optimize car evaluation classification. By systematically exploring 24 unique hyperparameter combinations, NeuroDrive identifies the peak-performance configuration for predicting vehicle marketability using the UCI Car Evaluation dataset.

---

🛠️ Project Overview

NeuroDrive functions as a systematic machine learning pipeline that automates the training and evaluation of multi-layer perceptron (MLP) models. The system performs rigorous data preprocessing, executes an exhaustive grid search across multiple architectural dimensions, and generates detailed visualizations to pinpoint model convergence patterns.

🧬 Technical Core

• Architectural Depth: Comparative analysis between 2-layer and 3-layer hidden networks.
• Optimization Logic: Benchmarking $logistic$, $tanh$, and $ReLU$ activation functions at varying learning rates ($0.01$ and $0.1$).
• Visual Intelligence: Automatically generates loss curves, accuracy heatmaps, and error comparisons for all configurations.
• Data Integrity: Implements a robust preprocessing workflow including one-hot encoding, duplicate removal, and feature standardization via StandardScaler.

---

📊 Dataset & Preprocessing

The Car Evaluation Dataset (via ucimlrepo) contains categorical attributes describing vehicles, including buying price, maintenance cost, safety ratings, and capacity.

To prepare the data for the neural network, the following pipeline is executed:

1. Cleaning: Handling missing values via dropna() and removing duplicates.
2. Encoding: Converting categorical variables into numerical format using One-Hot Encoding.
3. Scaling: Standardizing features using StandardScaler to ensure uniform gradient descent.
4. Partitioning: Splitting data into Training (80%) and Testing (20%) sets.

---

⚙️ Model Configurations

Using Scikit-learn's MLPClassifier, the suite evaluates 24 distinct models based on the following hyperparameter grid:

Hyperparameter	Values Tested
Activation Functions	logistic, tanh, relu
Learning Rates	0.01, 0.1
Training Epochs	100, 200
Hidden Layers	2 Layers, 3 Layers

Evaluation Metrics

Each configuration is benchmarked using:

• Accuracy: Training vs. Test classification scores.
• Loss/Error: Mean Squared Error (MSE) and loss history per epoch.

---

📁 Output & Analytics

The program generates an outputs/ directory containing the following artifacts:

Result Tables

• neural_network_results.csv: Raw data for all 24 model configurations.
• best_models.csv: A filtered list of models ranked by top test accuracy.

Performance Plots

• loss_vs_epochs.png: Convergence trajectories for all models.
• accuracy_heatmap.png: High-level visualization of hyperparameter impact on accuracy.
• error_comparison.png: Comparison of training vs. testing MSE.

---

🚀 Getting Started

Prerequisites

pip install pandas numpy matplotlib seaborn scikit-learn ucimlrepo

Execution

python neural_network.py

Note: The program will automatically iterate through all combinations and save results to the /outputs folder.

---

📝 Notes

• This project utilizes the Scikit-learn MLPClassifier for all neural network implementations.
• The benchmarking logic is modular; the code can be easily modified in neural_network.py to include additional layers or different solver types (e.g., adam, sgd).