Computer Science Department
School of Computer Science, Carnegie Mellon University
Goal-Directed Acting with Incomplete Information
We demonstrate how to use Partially Observable Markov Decision Process (POMDP) models to act, plan, and learn despite the uncertainty that results from actuator and sensor noise and missing information about the environment. We show how to use exponential utility functions to act in the presence of deadlines or in high-risk situations and demonstrate how to perform representation changes that transform planning tasks with exponential utility functions to planning tasks that standard search and planning methods from artificial intelligence can solve. Finally, we show how to decrease the planning time by interleaving planning and plan execution and present a real-time search method that allows for fine-grained control over how much planning to do between plan executions, uses heuristic knowledge to guide planning, and improves its performance over time as it solves similar planning tasks.
We use goal-directed robot-navigation tasks to illustrate the methods throughout the thesis, and present theoretical analyses, simulations, and experiments on a real robot.