Domain-independent Plan Intervention When Users Unwittingly Facilitate AttacksDownload PDF

Anonymous

Published: 24 May 2019, Last Modified: 05 May 2023XAIP 2019Readers: Everyone
Keywords: Plan/Goal recognition, machine learning, explainable planning
TL;DR: We introduce a machine learning model that uses domain-independent features to estimate the criticality of the current state to cause a known undesirable state.
Abstract: In competitive situations, agents may take actions to achieve their goals that unwittingly facilitate an opponent’s goals. We consider a domain where three agents operate: (1) a user (human), (2) an attacker (human or a software) agent and (3) an observer (a software) agent. The user and the attacker compete to achieve different goals. When there is a disparity in the domain knowledge the user and the attacker possess, the attacker may use the user’s unfamiliarity with the domain to its advantage and further its own goal. In this situation, the observer, whose goal is to support the user may need to intervene, and this intervention needs to occur online, on-time and be accurate. We formalize the online plan intervention problem and propose a solution that uses a decision tree classifier to identify intervention points in situations where agents unwittingly facilitate an opponent’s goal. We trained a classifier using domain-independent features extracted from the observer’s decision space to evaluate the “criticality” of the current state. The trained model is then used in an online setting on IPC benchmarks to identify observations that warrant intervention. Our contributions lay a foundation for further work in the area of deciding when to intervene.
4 Replies

Loading