Evolutionary algorithm creates lightweight CNN for real-time autonomous steering
Researchers shrink CNN models by 40% using a 1/5 success rule evolution strategy
A team of researchers led by Devson Butani at Lawrence Technological University (LTU) has applied an (N+M) Evolution Strategy (ES) with the classic 1/5th success rule to automatically optimize the hyperparameters of convolutional neural networks (CNNs) for autonomous steering. The goal was to find a lightweight network capable of predicting steering angles in real time, mimicking human driving behavior on predefined paths. The ES algorithm dynamically tuned filter sizes, layer configurations, and other parameters, effectively trimming unnecessary complexity while maintaining competitive accuracy.
The training data came from the LTU ACTor autonomous driving platform, capturing timestamped images with steering angle labels across varied driving styles and path directions. Despite the very small dataset, the optimized models achieved strong performance. The resulting lightweight CNNs are ideal for deployment on low-power, cost-effective hardware, striking a balance between computational efficiency and steering accuracy. This work paves the way for scalable, real-time autonomous driving systems without requiring massive neural networks or expensive GPUs.
- Used (N+M) Evolution Strategy with 1/5th success rule for hyperparameter optimization
- Reduced CNN size significantly while maintaining steering angle prediction accuracy
- Trained on small dataset from LTU ACTor platform, enabling real-time inference on low-cost hardware
Why It Matters
Enables cost-effective autonomous steering by replacing heavy CNNs with light, evolution-optimized models running on edge devices.