DISSERTATION DEFENSE
Reflective, Deliberative Agent-Based Information Gathering
Adam Eck
Dr. Leen-Kiat Soh (Advisor)
Dr. Ashok Samal, Dr. Stephen Scott, Dr. Steven Dunbar, and Dr. Milind Tambe
Thursday, April 23, 2015; 11:00am
347 Avery Hall
Abstract:
As computational devices and entities become further established as routine, omnipresent components of our everyday lives (e.g., wearable sensors, smart homes, cyber-physical systems, embodied agents, human-robot interactions), such systems face an increased pressure to perpetually understand the complex, noisy, uncertain world around them in real-time. This environmental knowledge enables computational systems to intelligently decide how to best behave in response to the current situation, adapt to the ever-changing conditions of the dynamic world, and accomplish system goals that ultimately aim to improve our daily experience. However, achieving and maintaining such knowledge is very complicated due to the complexities and challenging properties of real-world environments.
In this research, we study how to improve environment knowledge in intelligent agents and multiagent systems through reflective, deliberative information gathering. By being deliberative, an agent intentionally and selectively chooses how to gather information. By being reflective, an agent can self-evaluate its informational needs and performance in order to understand its needs and past sensing outcomes to best guide deliberative information gathering, as well as adapt and learn in an uncertain environment.
Within reflective, deliberative information gathering, this dissertation addresses two key problems: (1) the Analysis Problem, whereby an agent must determine how to measure and balance sensing benefits and costs in order to reflect and improve deliberative information gathering, (2) the Information Sharing Problem, whereby multiple agents must determine how to cooperatively sense together and share information to update collective beliefs.
For the Analysis Problem, we propose two improvements to a popular framework for reasoning under uncertainty—partially observable Markov decision processes (POMDPs): (1) Potential-based Reward Shaping (PBRS) providing metareasoning about information gathering within time-constrained planning, and (2) Difference-based Heuristic Selection (DHS) with Long Sequence Entropy Minimization (LSEM) for situationally-aware planning capable of balancing knowledge improvement and costs minimization. For the Information Sharing Problem, we propose two solutions for improving large team information sharing observing localized, non-stationary phenomena: (3) cooperative change detection and response and (4) forgetting-based adaptation of information sharing. We also propose: (5) a learning-based approach for ad hoc information gathering that enables agents to learn how to share information without requiring pre-coordination.