Optimization-based Machine Learning forDynamic Decision Problems
Abstract
Over the past decade, Machine Learning (ML) has had tremendous success in significantly advancingthe state of the art in numerous areas such as pattern recognition, speech processing, computervision, Natural Language Processing, and related disciplines. Over the next decade, the biggestgenerator of data is expected to be devices which sense and control the physical world. This explosionof real-time data that is emerging from the physical world requires a rapprochement of areassuch as machine learning, model-based dynamical systems, and control and decision theory. Whilecontrol theory has been firmly rooted in tradition of model-based design, the availability and scaleof data (both temporal and spatial) will require rethinking of the foundations for dynamic decisionproblems. One ofthe main challenges facing the Navy going forward will be to provide a rigorousfoundation for ML in the context of data-driven decision making and control and learning-basedoptimization of dynamical processes. However, most applications of ML to control lack the typeof guarantees that applications like autonomy demand. We further note that there is a sizable gapbetween mathematical theory of optimization, and the practice of optimization in machine learning.In this proposal, the two PIs, who have a history of a decade-long, fruitful collaborationmade possible by prior ONR support, propose to address both of these challenges and to expandthe methodological toolkit of Machine Learning to make it applicable for control of safety-criticalcomplex processes, and to further fill in the theory-practice gap in using optimization theory formachine learning. Our proposed approach is organized around three thrusts. The first t hrust iscentered around data driven, optimization-based learning of dynamical systems. In the secondthrust we will focus on advancing optimization based control and Reinforcement Learning (RL),from both a control-theoretic and online learning perspectives. In the third thrust, we will focuson utilizing statistical and computational advantages and trade offs in exploiting a particular notionof low-rank functions to provably speed up learning and to further close the gap between thetheory and practice of machine learning. In particular, we will attempt to create an explanationfor experimental observations in machine learning practice that seem to violate key assumptionsin optimization. Along the same lines, we will provide an explanation for why the phenomenon ofoversmoothing occurs in graph neural networks (GNNs), which have become increasingly popularin Machine learning practice for learning from data that has a graph structure.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2023
- Source ID
- N000142312299
Entities
People
- Ali Jadbabaie
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy