Making and Keeping Probabilistic Commitments for Trustworthy Multiagent Coordination

Abstract

In a large number of real world domains, such as the control of autonomous vehicles, team sports, medical diagnosis and treatment, and many others, multiple autonomous agents need to take actions based on local observations, and are interdependent in the sense that they rely on each other to accomplish tasks. Thus, achieving desired outcomes in these domains requires interagent coordination. The form of coordination this thesis focuses on is commitments, where an agent, referred to as the commitment provider, specifies guarantees about its behavior to another, referred to as the commitment recipient, so that the recipient can plan and execute accordingly without taking into account the details of the provider's behavior. This thesis grounds the concept of commitments into decision-theoretic settings where the provider's guarantees might have to be probabilistic when its actions have stochastic outcomes and it expects to reduce its uncertainty about the environment during execution.

More concretely, this thesis presents a set of contributions that address three core issues for commitment-based coordination: probabilistic commitment adherence, interpretation, and formulation. The first contribution is a principled semantics for the provider to exercise maximal autonomy that responds to evolving knowledge about the environment without violating its probabilistic commitment, along with a family of algorithms for the provider to construct policies that provably respect the semantics and make explicit tradeoffs between computation cost and plan quality. The second contribution consists of theoretical analyses and empirical studies that improve our understanding of the recipient's interpretation of the partial information specified in a probabilistic commitment; the thesis shows that it is inherently easier for the recipient to robustly model a probabilistic commitment where the provider promises to enable preconditions that the recipient requires than where the provider instead promises to avoid changing already-enabled preconditions. The third contribution focuses on the problem of formulating probabilistic commitments for the fully cooperative provider and recipient; the thesis proves structural properties of the agents' values as functions of the parameters of the commitment specification that can be exploited to achieve orders of magnitude less computation for 1) formulating optimal commitments in a centralized manner, and 2) formulating (approximately) optimal queries that induce (approximately) optimal commitments for the decentralized setting in which information relevant to optimization is distributed among the agents.