Autonomous driving solutions

Modern autonomous vehicles are powered by reinforcement learning based command and control architectures. These architectures are augmented with insights and analyses gleaned from auxiliary machine learning platforms, that operate on data collected from sensors, thus together driving innovations aimed at mitigating risk, enhancing safety, and improving real-time experience.

Several issues needed to be addressed in order to optimize the entire pipeline for our client. These included data preprocessing, cleansing; data cleansing and feature selection; contextualizing; specifying the objectives and constraints (e.g., inference time, hardware imposed restrictions) and outputs; formulating appropriate models such as POMDP and incorporating incremental partial (bandit) feedback, designing and implementing appropriate efficient and tractable (approximate) sampling algorithms; pretraining (since collecting the data is typically really small and expensive, and leveraging other non-idealistic but larger and less expensive data sources becomes critical); off-policy evaluation and optimization; uncertainty quantification; and hyperparameter optimization (using AutoML).