• Sorted by Date • Classified by Publication Type • Classified by Topic • Grouped by Student (current) • Grouped by Former Students •
Mor Vered. Mirroring: A General Approach to Plan and Goal
Recognition. Ph.D. Thesis, Bar Ilan University, 2018.
Plan recognition is the task of inferring the plan of an agent, based on an incomplete sequence of its observed actions. The problem may be further sub categorized into offline and online plan recognition. In offline versions ofthe problem, the entire sequence is given to the recognizer at once. In contrast, in online recognition the observations are provided incrementally. The traditional approach to plan recognition has been to compare observations against a dedicated plan library representing all known plans to achieve known goals in a manner that facilitates efficient inference. This approach fails when encountering unknown plans or when dealing with continuous domains where the potential plan possibilities may be infinite. A recent approach to plan recognition uses a planner to dynamically generate plans for given goals, thus eliminates the need for the a traditional plan library. While an inspiring approach it poses many problems for online recognition in continuous environments.A key problem in previous formulations of plan recognition over continuous domains is the early commitment to specific discretizations of the environment and the observed agent's actions, often leading to a reduction in recognition accuracy. To address this we introduce Mirroring. Inspired by mirroring processes hypothesized to take place in human brains, Mirroring is a formalization of recognition problems which admits continuous environments, as well as discrete domains. We further develop Mirroring as a complete online goal recognition approach that uses a black-box planner to generate recognition hypotheses, avoiding the prevalent assumption in current approaches, which rely on a dedicated plan library. Due to the suitability over continuous domains we can now apply continuous-world motion planners in plan recognition. % Such methods are inherently limited to the knowledge represented in the library. ---we can apply continuous-world motion planners in plan recognition. We proceed to provide formal arguments for the usefulness of Mirroring, and empirically evaluate Mirroring over a thousand recognition problems in three continuous domains and six classical planning domains. We also proceed to contrast machine and human recognition in two challenging domains, revealing insights as to human capabilities; and finally we compare Mirroring to library-based methods. As Mirroring requires multiple calls to a planner within the recognition process it can be inefficient for online recognition.Recognizing goals with minimal domain knowledge as an agent executes its plan requires efficient algorithms to sift through a large space of hypotheses. We therefore identify two independent decision points within the Mirroring algorithm where heuristicsmay be used to improve online run-time. We specify such heuristics for continuous domains, prove guarantees on their use, and empirically evaluate both the performance and efficiency of our algorithm over hundreds of experiments in both a 3D navigational environment and a cooperative robotic team task. We additionally test the durability of our approach by experimenting over scenarios with varying recognition difficulty, with both evenly spread and clustered goals. As a final optimization method we further develop an online approach to recognize goals in both continuous and discrete domains using a combination of Mirroring and a generalized notion of landmarks adapted from the planning literature. Extensive experiments demonstrate the approach is more efficient and substantially more accurate than state-of-the-art.
@PhdThesis{mor-phd,
author = {Mor Vered},
title = {Mirroring: A General Approach to Plan and Goal Recognition},
school = {{B}ar {I}lan {U}niversity},
year = {2018},
wwwnote = {},
OPTannote = {} ,
abstract = {
Plan recognition is the task of inferring the plan of an agent, based on an incomplete sequence of its observed actions. The problem may be further sub categorized into offline and online plan recognition. In \textit{offline} versions of
the problem, the entire sequence is given to the recognizer at once. In contrast, in \textit{online}
recognition the observations are provided incrementally. The traditional approach to plan recognition has been to compare observations against a dedicated \emph{plan library} representing all known plans to achieve known goals in a manner that facilitates efficient inference. This approach fails when encountering unknown plans or when dealing with continuous domains where the potential plan possibilities may be infinite.
A recent approach to plan recognition uses
a planner to dynamically generate plans for given goals, thus eliminates the need for the a traditional plan library. While an inspiring approach it poses many problems for online recognition in continuous environments.
A key problem in previous formulations of plan recognition over \emph{continuous} domains is the early commitment to specific discretizations of the environment and the observed agent's actions, often leading to a reduction in recognition accuracy.
To address this we introduce \textit{Mirroring}. Inspired by mirroring processes hypothesized to take place in human brains, \emph{Mirroring} is a formalization of recognition problems which admits continuous environments, as well as discrete domains.
We further develop \textit{Mirroring} as a complete online goal recognition approach that uses a black-box planner to generate recognition hypotheses, avoiding the prevalent assumption in current approaches, which rely on a dedicated \emph{plan library}. Due to the suitability over continuous domains we can now apply continuous-world motion planners in plan recognition. % Such methods are inherently limited to the knowledge represented in the library. ---we can apply continuous-world motion planners in plan recognition.
We proceed to provide formal arguments for the usefulness of Mirroring, and empirically evaluate Mirroring over a thousand recognition problems in three continuous domains and six classical planning domains. We also proceed to contrast machine and human recognition in two challenging domains, revealing insights as to human capabilities; and finally we compare Mirroring to library-based methods.
As Mirroring requires multiple calls to a planner within the recognition process it can be inefficient for online recognition.
Recognizing goals with minimal domain knowledge as an agent executes its plan requires efficient algorithms to sift through a large space of hypotheses.
We therefore identify two independent decision points within the \emph{Mirroring} algorithm where heuristics
may be used to improve online run-time. We specify such heuristics for continuous domains, prove guarantees on their use, and empirically evaluate both the performance and efficiency of our algorithm over hundreds of experiments in both a 3D navigational environment and a cooperative robotic team task. We additionally test the durability of our approach by experimenting over scenarios with varying recognition difficulty, with both evenly spread and clustered goals.
As a final optimization method we further develop an online approach to recognize goals in both continuous and discrete domains using a combination of \emph{Mirroring} and a generalized notion of landmarks adapted from the planning literature. Extensive experiments demonstrate the approach is more efficient and substantially more accurate than state-of-the-art.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Sun Feb 08, 2026 18:23:11