# Grid Search and Genetic Algorithm Optimization of Neural Networks for Automotive Radar Object Classification

## TL;DR

This paper compares grid search and genetic algorithm optimization for neural networks in classifying objects in automotive radar systems, finding that genetic algorithms yield higher accuracy.

## Contribution

A novel comparison of grid search and genetic algorithm optimization for radar object classification, demonstrating improved accuracy with GA-optimized deep neural networks.

## Key findings

- GA-optimized model achieves 97.40% mean accuracy using expanded radar features and deeper architecture.
- Grid search model reaches 90.06% validation accuracy with a compact neural network and motion descriptors.
- PCA and SHAP analysis confirms the enhanced discriminative power of the expanded feature set.

## Abstract

This paper proposes and evaluates two neural network-based approaches for object classification in automotive radar systems, comparing the performance impact of grid search and genetic algorithm (GA) hyperparameter optimization strategies. The task involves classifying cars, pedestrians, and cyclists using radar-derived features. The grid search–optimized model employs a compact architecture with two hidden layers and 10 neurons per layer, leveraging kinematic correlations and motion descriptors to achieve mean accuracies of 90.06% (validation) and 90.00% (test). In contrast, the GA-optimized model adopts a deeper architecture with nine hidden layers and 30 neurons per layer, integrating an expanded feature set that includes object dimensions, signal-to-noise ratio (SNR), radar cross-section (RCS), and Kalman filter–based motion descriptors, resulting in substantially higher performance at approximately 97.40% mean accuracy on both validation and test datasets. Principal Component Analysis (PCA) and SHapley Additive exPlanations (SHAP) highlight the enhanced discriminative power of the new set of features, while parallelized GA execution enables efficient exploration of a broader hyperparameter space. Although currently optimized for urban traffic scenarios, the proposed approach can be extended to highway and extra-urban environments through targeted dataset expansion and developing additional features that are less sensitive to object kinematics, thereby improving robustness across diverse motion patterns and operational contexts.