Tensor Data Compression and Dimensionality Reduction for Autonomous Mobility

Project #1.32 duration 2020-2021.

High-fidelity, physics-based simulation has become an indispensable part of learning and cognitive pipelines for autonomous vehicles in off-road settings; it enables offline modeling and testing of a system without incurring the cost and risk factors associated with real-life experimentation. In high-fidelity soil-vehicle models, simulations are not only computationally expensive, a pervasive challenge in applications to autonomy arises from the storage, analysis and maintenance of extremely large data sets. Current state-of-the-art setups keep the bare minimum of the computational results that are necessary for the application at hand, and discard the rest. This precludes any knowledge that may be obtained from the entire data-set.

The primary objective of our research is to use tensor decomposition methods to construct data-driven, reduced-order surrogate models for expensive, full-scale physics-based models and propose a cost-aware multi-fidelity active learning framework. By combining our generated multi-fidelity models with the proposed learning framework, we aim to accelerate learning tasks relevant to autonomous mobility.

In our approach, high-fidelity discrete element method (DEM) simulation data is compressed using tensor decomposition methods. This allows us to keep a highly compressed version of the entire data set which can be fully exploited to train learning algorithms, produce enhanced reduced-order models, and perform offline-online learning by providing a compressed model which can be efficiently stored and updated by computing resources on board the vehicle for learning tasks in an online setting.

Multi-fidelity modeling has been widely used to enhance parametric space exploration and learning outcomes when, as is the case for DEM simulations, our high-fidelity model must be queried sparingly due to its computational cost. We used our tensor compression approach to construct a multi-fidelity sequence of surrogate models. Then, we employed this sequence of models within a novel cost-aware active learning framework with multi-fidelity sources, in which the model providing the most information gain per unit of computational cost is queried to continue exploring the parametric space.

1.32