Model Fusion: Weighted N-Version Programming for Resilient Autonomous Vehicle Steering Control

Abstract

We present the preliminary results on developing a weighted N-version programming (NVP) scheme for ensuring resilience of machine learning based steering control algorithms. The proposed scheme is designed based on the fusion of outputs from three redundant Deep Neural Network (DNN) models, independently designed using Udacity's self driving car challenge data. The improvement in reliability compared to single DNN models is evaluated by measuring the steering angle prediction accuracy in the presence of simulated perturbations on input image data caused by various environmental conditions.