TBD
Abstract
The fields of artificial intelligence (AI) and machine learning (ML) have had significant impact in recent years, and yet it is increasingly clear that further progress is being hampered by a narrow focus on pattern recognition, and on replicating the intelligence displayed by a single human, in a limited class of decision-making tasks. The natural scope of learning systems is much broader than single-node intelligence—it is the real-world socio-technical networks of planetary scope that are currently emerging in domains such as commerce, transportation, healthcare, and defense. Currently AI/ML methods are merely being dropped into existing networks in an adhoc manner. This is a major intellectual mismatch. The intelligence supplied at the network level by microeconomic mechanisms is complemen tary to the kind of intelligence sought in the classical conception of AI. Specifically, markets and game-theoretic mechanisms can be adaptive, robust, long-lasting, and scalable—all clearly desiderata of intelligent systems. Moreover, if suitably regulated they can be fair and transparent This research project focuses on the development of a new bridging discipline that can be construed as “learning-aware microeconomics” or as “market-aware machine learning.” Specific concrete projects that illustrate this agenda and highlight the underlying mathematical challenges are as follows: (1) multi-arm bandit learning in multi-way matching markets, where agents learn to choose actions based on observed data while managing an exploration/exploitation tradeoff and where agents are in competition over a scarce resource across a network; (2) a new notion of a “data economy” across a network of agents, where a blend of statistical and inferential concepts gives rise to a rise to an economic valuation of a data resource, and where an auction mechanism provides incentives for sharing data; and (3) a revolutionary new approach to constrained optimization, based on the variational mechanical concepts of virtual work and virtual power, providing a flexible way to interface dynamical, economic, and inferential subsystems. The anticipated outcome of the successful completion of the research proposed here is a sub stantial expansion of the current research agenda in statistical machine learning to incorporate microeconomic perspectives, and the training of a cohort of researchers prepared to pursue that agenda for many years to come. We anticipate new design and analysis procedures for algorithms that learn from data in settings involving scarce resources, where those resources are valued by agents as part of the learning procedure, and where the scarce resources may themselves be data. Networks play a fundamental role in our conception of this research agenda—not merely net works as observed data or networks as modeled solely by probability laws, but rather networks as flows of economic and informational value, and as coordinators of interacting decision-making under uncertainty. The anticipated DoD benefits involve new capabilities in the design, analysis, and deployment of information-and-control systems: (1) actionable consequences from large-scale data streams; (2) decision-oriented machine learning in a decentralized setting; (3) design principles for robust learning systems; (4) management of exploration-exploitation tradeoffs across networks of self interested agents, including adversaries; and (5) guidance for realizing tradeoffs between compu tational, inferential, and game-theoretic criteri
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 09, 2022
- Source ID
- N000142112941
Entities
People
- Michael I. Jordan
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents