Helping humans and agents avoid undesirable consequences with models of intervention
Date
2021
Authors
Weerawardhana, Sachini Situmini, author
Whitley, Darrell, advisor
Ray, Indrajit, committee member
Pallickara, Sangmi, committee member
Ortega, Francisco, committee member
Seger, Carol, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
When working in an unfamiliar online environment, it can be helpful to have an observer that can intervene and guide a user toward a desirable outcome while avoiding undesirable outcomes or frustration. The Intervention Problem is deciding when to intervene in order to help a user. The Intervention Problem is similar to, but distinct from, Plan Recognition because the observer must not only recognize the intended goals of a user but also when to intervene to help the user when necessary. In this dissertation, we formalize a family of intervention problems to address two sub-problems: (1) The Intervention Recognition Problem, and (2) The Intervention Recovery Problem. The Intervention Recognition Problem views the environment as a state transition system where an agent (or a human user), in order to achieve a desirable outcome, executes actions that change the environment from one state to the next. Some states in the environment are undesirable and the user does not have the ability to recognize them and the intervening agent wants to help the user in the environment avoid the undesirable state. In this dissertation, we model the environment as a classical planning problem and discuss three intervention models to address the Intervention Recognition Problem. The three models address different dimensions of the Intervention Recognition Problem, specifically the actors in the environment, information hidden from the intervening agent, type of observations and noise in the observations. The first model: Intervention by Recognizing Actions Enabling Multiple Undesirable Consequences, is motivated by a study where we observed how home computer users practice cyber-security and take action to unwittingly put their online safety at risk. The model is defined for an environment where three agents: the user, the attacker and the intervening agent are present. The intervening agent helps the user reach a desirable goal that is hidden from the intervening agent by recognizing critical actions that enable multiple undesirable consequences. We view the problem of recognizing critical actions as a multi-factor decision problem of three domain-independent metrics: certainty, timeliness and desirability. The three metrics simulate the trade-off between the safety and freedom of the observed agent when selecting critical actions to intervene. The second model: Intervention as Classical Planning, we model scenarios where the intervening agent observes a user and a competitor attempting to achieve different goals in the same environment. A key difference in this model compared to the first model is that the intervening agent is aware of the user's desirable goal and the undesirable state. The intervening agent exploits the classical planning representation of the environment and uses automated planning to project the possible outcomes in the environment exactly and approximately. To recognize when intervention is required, the observer analyzes the plan suffixes leading to the user's desirable goal and the undesirable state and learns the differences between the plans that achieve the desirable goal and plans that achieve the undesirable state using machine learning. Similar to the first model, learning the differences between the safe and unsafe plans allows the intervening agent to balance specific actions with those that are necessary for the user to allow some freedom. The third model: Human-aware Intervention, we assume that the user is a human solving a cognitively engaging planning task. When human users plan, unlike an automated planner, they do not have the ability to use heuristics to search for the best solution. They often make mistakes and spend time exploring the search space of the planning problem. The complication this adds to the Intervention Recognition Problem is that deciding to intervene by analyzing plan suffixes generated by an automated planner is no longer feasible. Using a cognitively engaging puzzle solving task (Rush Hour) we study how human users solve the puzzle as a planning task and develop the Human-aware Intervention model combining automated planning and machine learning. The intervening agent uses a domain specific feature set more appropriate for human behavior to decide in real time whether to intervene the human user. Our experiments using the benchmark planning domains and human subject studies show that the three intervention recognition models out performs existing plan recognition algorithms in predicting when intervention is required. Our solution to address the Intervention Recovery Problem goes beyond the typical preventative measures to help the human user recover from intervention. We propose the Interactive Human-aware Intervention where a human user solves a cognitively engaging planning task with the assistance of an agent that implements the Human-aware Intervention. The Interactive Human-aware Intervention is different from typical preventive measures where the agent executes actions to modify the domain such that the undesirable plan can not progress (e.g., block an action). Our approach interactively guides the human user toward the solution to the planning task by revealing information about the remaining planning task. We evaluate the Interactive Human-aware Intervention using both subjective and objective measures in a human subject study.
Description
Rights Access
Subject
automated planning
human-aware planning
artificial intelligence
plan/goal recognition
cyber-security