Accelerating Adaptivity under Limited Data and Computation: A Meta-Learning Approach
Abstract
This research program is motivated by our long-term vision to understand how to reduce the sample complexity required to achieve learning in autonomous systems in data-starved or non-stationary domains. Indeed, traditional learning tools necessitate the availability of massive labeled data sampled from static distributions to find rich models with plausible performance. Doing so is inapplicable to the battlefield, mostly because acquiring such data is impractically hazardous or time-consuming. To enable learning with limited data, Meta-Learning provides a framework for adapting to unseen tasks by exploiting their similarity to seen tasks. A particularly successful form of Meta-Learning is Model-Agnostic Meta-Learning (MAML), which does not require any assumption on the common structure between tasks and simply tries to come up with a proper initial model that can be quickly adapted to new tasks with a few steps of gradient-based updates. Although MAML has been very successful in practices such as robotics and image classification, several theoretical aspects of this framework, such as its generalization guarantees, performance in dynamic environments, extensions to settings with discrete variables, and application to multi-user (distributed) settings are not yet well-studied. This proposal describes a research program to develop novel and foundational frameworks to study these problems. The proposed effort is divided into three thrusts. (i) The first thrust studies MAML in dynamic environments where the task probability distribution changes, and, in particular, it focuses on Meta-Reinforcement Learning. (ii) The current theory and application of MAML is limited to continuous settings, and the second thrust tries to provide a novel extension of this framework to discrete settings. (iii) The third thrust focuses on the application of MAML in multi-agent settings to obtain individualized, adaptable models for networks of autonomous agents. These research thrusts will be considered in the context of data acquired from ground robots, and will alleviate the statistical limitations of applying learning tools to this domain.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110226
Entities
People
- Aryan Mokhtari
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Austin