Goal recognition and deception in path-planning

Masters, P 2019, Goal recognition and deception in path-planning, Doctor of Philosophy (PhD), Science, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Masters.pdf Thesis application/pdf 2.82MB
Title Goal recognition and deception in path-planning
Author(s) Masters, P
Year 2019
Abstract This thesis argues that investigation of goal recognition and deception in the much studied and well-understood context of path-planning reveals nuances to both problems that have previously gone unnoticed. Contemporary goal recognition systems rely on examination of multiple observations to calculate a probability distribution across goals. The first part of this thesis demonstrates that a distribution with identical rankings to current stateof-the-art can be achieved without any observations apart from a known starting point (such as a door or gate) and where the agent is now. It also presents a closed formula to calculate a radius around any goal of interest within which that goal is guaranteed to be the most probable, without having to calculate any actual probability values. In terms of deception, traditionally there are two strategies: dissimulation (hiding the true) and simulation (showing the false). The second part of this thesis shows that current state-of-the-art goal recognition systems do not cope well with dissimulation that does its work by ‘dazzling’ (i.e., obfuscating with hugely suboptimal plans). It presents an alternative, self-modulating formula that modifies its output when it encounters suboptimality, seeming to ‘know that it does not know’ instead of ‘keep changing its mind’. Deception is often regarded as a ‘yes, no’ proposition (either the target is deceived or they are not). Furthermore, intuitively, deceptive path-planning involves suboptimality and must, therefore, be expensive. This thesis, however, presents a model of deception for path-planning domains within which it is possible (a) to rank paths by their potential to deceive and (b) to generate deceptive paths that are ‘optimally deceptive’ (i.e., deceptive to the maximum extent at the lowest cost).
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Science
Subjects Adaptive Agents and Intelligent Robotics
Keyword(s) deception
goal recognition
planning
path-planning
adversarial reasoning
Versions
Version Filter Type
Access Statistics: 48 Abstract Views, 63 File Downloads  -  Detailed Statistics
Created: Thu, 30 May 2019, 16:35:27 EST by Keely Chapman
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us