NICOP - Large Scale Sequential Decision Making in An Uncertain World
Abstract
Given the current abundance and variety of sensors and data, making automated decisions under uncertainty is as relevant and difficult as it has ever been. These problems are partially observable sequential decision problems with huge number of state, observation, and action variables. Par- tially Observable Markov Decision Processes (POMDPs) have been developed to frame and solve such problems prudently, in a systematic and principled manner. However, although significant advances have been made in solving POMDP problems with large number of state and observation variables, problems with large number of action variables remains extremely challenging.This project will design fundamental algorithms and practical software tools to alleviate such chal- lenges. In particular, this project proposes a one year (with one additional option year) extensive study of Monte Carlo methods for automated decision making under uncertainty, formalized as POMDPs. It aims to design fundamental algorithms for solving large scale POMDP problems, es- pecially problems with huge number of action variables, by building on recent advances in Monte Carlo methods and stochastic optimization techniques. This will result in software tools for au- tonomous decision making that can leverage massive amount of sensors and data in real-time.This project will contribute to focus area Autonomy and Unmanned Systems of the Naval Science and Technology Strategy, published in February 2015. It will support the focus area~s objective on Intelligence Enablers and Architectures in Scalable planning and re-planning. Furthermore, this project is aligned with two areas of interest of Code 31, Division 311, Machine Learning, Reasoning, and Intelligence Program, thrust Intelligence for Autonomous Agents, which are:~ Planning in large domains in partially known environments and incompletely modeled goals and domains.~ Intelligent architectures that seamlessly integrate knowledge-bases, learning, reasoning, and planning, for decision-making.Desired outcomes (per year) of this research project are:~ A new algorithm for solving decision making problems framed as POMDPs.~ A conference and a journal papers in an international conference and journal venues for Artificial Intelligence or Robotics, such as Int. Conference on Automated Planning and Scheduling (ICAPS) and ITEE Trans. on Robotics.~ A software implementation of the new method.This proposal is an extension of the pre-proposal and white paper discussions with Drs. Hiekeun~Higgin~ Ko and Jason Wong.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N629091712046
Entities
People
- Hanna Kurniawati
Organizations
- Office of Naval Research
- United States Navy
- University of Queensland