Corresponding Author: Sumedha Dangi
Deep learning pipelines are essential in perception, planning, and control of autonomous vehicles (AVs). The optimization approach, defining the efficiency of models in learning and generalizing in dynamic environments, is an essential but under-researched aspect of such pipelines. Current methods, such as SGD, Adam, and variants like BAAO and GASGD, do not change during the training process and cannot adapt to the various real-world conditions. Here, Meta-Optimization Framework to Fault Prediction (MOCAFP) is introduced, a kind of adaptive controller, which uses the gradient variance and convergence rate as training inputs to choose and optimize optimizers dynamically. The system was evaluated on four datasets that consisted of simulated and real-world driving conditions: Udacity Jungle, Udacity Lake, KITTI, and a real-world Urban Roads dataset. Experiments show that MOCAFP shows better performance than baseline optimizers, in terms of convergence speed and learning stability. In particular, it is able to attain competitive training times and improve the coefficient of determination (R2) by 12 per cent, reduce prediction variance by 15 per cent, and achieve a minimum mean absolute error (MAE) of up to 18 per cent. The improvements imply that MOCAFP is a reliable and scalable approach towards the future of fault prediction in AV pipes.