Synthesis of Interpretable and Obfuscatory Behaviors in Human-Aware AI Systems

161301-Thumbnail Image.png
Description
In settings where a human and an embodied AI (artificially intelligent) agent coexist, the AI agent has to be capable of reasoning with the human's preconceived notions about the environment as well as with the human's perception limitations. In addition,

In settings where a human and an embodied AI (artificially intelligent) agent coexist, the AI agent has to be capable of reasoning with the human's preconceived notions about the environment as well as with the human's perception limitations. In addition, it should be capable of communicating intentions and objectives effectively to the human-in-the-loop. While acting in the presence of human observers, the AI agent can synthesize interpretable behaviors like explicable, legible, and assistive behaviors by accounting for the human's mental model (inclusive of her sensor model) in its reasoning process. This thesis will study different behavior synthesis algorithms which focus on improving the interpretability of the agent's behavior in the presence of a human observer. Further, this thesis will study how environment redesign strategies can be leveraged to improve the overall interpretability of the agent's behavior. At times, the agent's environment may also consist of purely adversarial entities or mixed entities (i.e. adversarial as well as cooperative entities), that are trying to infer information from the AI agent's behavior. In such settings, it is crucial for the agent to exhibit obfuscatory behavior that prevents sensitive information from falling into the hands of the adversarial entities. This thesis will show that it is possible to synthesize interpretable as well as obfuscatory behaviors using a single underlying algorithmic framework.
Date Created
2021
Agent

Planning challenges in human-robot teaming

153091-Thumbnail Image.png
Description
As robotic technology and its various uses grow steadily more complex and ubiquitous, humans are coming into increasing contact with robotic agents. A large portion of such contact is cooperative interaction, where both humans and robots are required to work

As robotic technology and its various uses grow steadily more complex and ubiquitous, humans are coming into increasing contact with robotic agents. A large portion of such contact is cooperative interaction, where both humans and robots are required to work on the same application towards achieving common goals. These application scenarios are characterized by a need to leverage the strengths of each agent as part of a unified team to reach those common goals. To ensure that the robotic agent is truly a contributing team-member, it must exhibit some degree of autonomy in achieving goals that have been delegated to it. Indeed, a significant portion of the utility of such human-robot teams derives from the delegation of goals to the robot, and autonomy on the part of the robot in achieving those goals. In order to be considered truly autonomous, the robot must be able to make its own plans to achieve the goals assigned to it, with only minimal direction and assistance from the human.

Automated planning provides the solution to this problem -- indeed, one of the main motivations that underpinned the beginnings of the field of automated planning was to provide planning support for Shakey the robot with the STRIPS system. For long, however, automated planners suffered from scalability issues that precluded their application to real world, real time robotic systems. Recent decades have seen a gradual abeyance of those issues, and fast planning systems are now the norm rather than the exception. However, some of these advances in speedup and scalability have been achieved by ignoring or abstracting out challenges that real world integrated robotic systems must confront.

In this work, the problem of planning for human-hobot teaming is introduced. The central idea -- the use of automated planning systems as mediators in such human-robot teaming scenarios -- and the main challenges inspired from real world scenarios that must be addressed in order to make such planning seamless are presented: (i) Goals which can be specified or changed at execution time, after the planning process has completed; (ii) Worlds and scenarios where the state changes dynamically while a previous plan is executing; (iii) Models that are incomplete and can be changed during execution; and (iv) Information about the human agent's plan and intentions that can be used for coordination. These challenges are compounded by the fact that the human-robot team must execute in an open world, rife with dynamic events and other agents; and in a manner that encourages the exchange of information between the human and the robot. As an answer to these challenges, implemented solutions and a fielded prototype that combines all of those solutions into one planning system are discussed. Results from running this prototype in real world scenarios are presented, and extensions to some of the solutions are offered as appropriate.
Date Created
2014
Agent

Planning with incomplete user preferences and domain models

152834-Thumbnail Image.png
Description
Current work in planning assumes that user preferences and/or domain dynamics are completely specified in advance, and aims to search for a single solution plan to satisfy these. In many real world scenarios, however, providing a complete specification of user

Current work in planning assumes that user preferences and/or domain dynamics are completely specified in advance, and aims to search for a single solution plan to satisfy these. In many real world scenarios, however, providing a complete specification of user preferences and domain dynamics becomes a time-consuming and error-prone task. More often than not, a user may provide no knowledge or at best partial knowledge of her preferences with respect to a desired plan. Similarly, a domain writer may only be able to determine certain parts, not all, of the model of some actions in a domain. Such modeling issues requires new concepts on what a solution should be, and novel techniques in solving the problem. When user preferences are incomplete, rather than presenting a single plan, the planner must instead provide a set of plans containing one or more plans that are similar to the one that the user prefers. This research first proposes the usage of different measures to capture the quality of such plan sets. These are domain-independent distance measures based on plan elements if no knowledge of the user preferences is given, or the Integrated Preference Function measure in case incomplete knowledge of such preferences is provided. It then investigates various heuristic approaches to generate plan sets in accordance with these measures, and presents empirical results demonstrating the promise of the methods. The second part of this research addresses planning problems with incomplete domain models, specifically those annotated with possible preconditions and effects of actions. It formalizes the notion of plan robustness capturing the probability of success for plans during execution. A method of assessing plan robustness based on the weighted model counting approach is proposed. Two approaches for synthesizing robust plans are introduced. The first one compiles the robust plan synthesis problems to the conformant probabilistic planning problems. The second approximates the robustness measure with lower and upper bounds, incorporating them into a stochastic local search for estimating distance heuristic to a goal state. The resulting planner outperforms a state-of-the-art planner that can handle incomplete domain models in both plan quality and planning time.
Date Created
2014
Agent

When is temporal planning really temporal

151471-Thumbnail Image.png
Description
In this dissertation I develop a deep theory of temporal planning well-suited to analyzing, understanding, and improving the state of the art implementations (as of 2012). At face-value the work is strictly theoretical; nonetheless its impact is entirely real and

In this dissertation I develop a deep theory of temporal planning well-suited to analyzing, understanding, and improving the state of the art implementations (as of 2012). At face-value the work is strictly theoretical; nonetheless its impact is entirely real and practical. The easiest portion of that impact to highlight concerns the notable improvements to the format of the temporal fragment of the International Planning Competitions (IPCs). Particularly: the theory I expound upon here is the primary cause of--and justification for--the altered (i) selection of benchmark problems, and (ii) notion of "winning temporal planner". For higher level motivation: robotics, web service composition, industrial manufacturing, business process management, cybersecurity, space exploration, deep ocean exploration, and logistics all benefit from applying domain-independent automated planning technique. Naturally, actually carrying out such case studies has much to offer. For example, we may extract the lesson that reasoning carefully about deadlines is rather crucial to planning in practice. More generally, effectively automating specifically temporal planning is well-motivated from applications. Entirely abstractly, the aim is to improve the theory of automated temporal planning by distilling from its practice. My thesis is that the key feature of computational interest is concurrency. To support, I demonstrate by way of compilation methods, worst-case counting arguments, and analysis of algorithmic properties such as completeness that the more immediately pressing computational obstacles (facing would-be temporal generalizations of classical planning systems) can be dealt with in theoretically efficient manner. So more accurately the technical contribution here is to demonstrate: The computationally significant obstacle to automated temporal planning that remains is just concurrency.
Date Created
2012
Agent

Partial satisfaction planning: representation and solving methods

151144-Thumbnail Image.png
Description
Automated planning problems classically involve finding a sequence of actions that transform an initial state to some state satisfying a conjunctive set of goals with no temporal constraints. But in many real-world problems, the best plan may involve satisfying only

Automated planning problems classically involve finding a sequence of actions that transform an initial state to some state satisfying a conjunctive set of goals with no temporal constraints. But in many real-world problems, the best plan may involve satisfying only a subset of goals or missing defined goal deadlines. For example, this may be required when goals are logically conflicting, or when there are time or cost constraints such that achieving all goals on time may be too expensive. In this case, goals and deadlines must be declared as soft. I call these partial satisfaction planning (PSP) problems. In this work, I focus on particular types of PSP problems, where goals are given a quantitative value based on whether (or when) they are achieved. The objective is to find a plan with the best quality. A first challenge is in finding adequate goal representations that capture common types of goal achievement rewards and costs. One popular representation is to give a single reward on each goal of a planning problem. I further expand on this approach by allowing users to directly introduce utility dependencies, providing for changes of goal achievement reward directly based on the goals a plan achieves. After, I introduce time-dependent goal costs, where a plan incurs penalty if it will achieve a goal past a specified deadline. To solve PSP problems with goal utility dependencies, I look at using state-of-the-art methodologies currently employed for classical planning problems involving heuristic search. In doing so, one faces the challenge of simultaneously determining the best set of goals and plan to achieve them. This is complicated by utility dependencies defined by a user and cost dependencies within the plan. To address this, I introduce a set of heuristics based on combinations using relaxed plans and integer programming formulations. Further, I explore an approach to improve search through learning techniques by using automatically generated state features to find new states from which to search. Finally, the investigation into handling time-dependent goal costs leads us to an improved search technique derived from observations based on solving discretized approximations of cost functions.
Date Created
2012
Agent