Faculty Recruiting Support CICS

Reliable Decision-Making with Imprecise Models

25 Jun
Thursday, 06/25/2020 1:00pm to 3:00pm
Zoom Meeting
PhD Dissertation Proposal Defense

Abstract:

The rapid growth in the deployment of autonomous systems in different fields has generated considerable interest in how agents can operate reliably in large, stochastic, and unstructured environments. In automated planning, agents reason about the effects of their actions based on the model parameters. Agents acting in the real-world, however, often operate based on models that do not represent all the details in the environment. The source of model imprecision ranges from model simplification for tractability to unawareness of missing information in the design phase. Reasoning with such imprecise models affects the solution quality and often results in undesirable agent behavior, some of which may be unsafe. The key question is how can we achieve well-behaved autonomous systems that can reason in the presence of uncertainty and model imprecision.

This thesis presents solutions to address three classes of model imprecision in a Markov decision process, along with an analysis of the conditions under which bounded-optimality can be achieved. A key challenge in improving the overall performance is to detect and measure the undesirable impacts of agent actions. The agent may not have prior knowledge of which actions are unsafe and the details about the side effects of its actions. This information is gathered through various forms of feedback, which the agent uses to update its policy for execution. Besides model imprecision, the reliability of decisions in the real-world may be affected by interpretability of the generated solutions. The impact of generating interpretable solutions is studied in the context of graph clustering for decision-support systems.

Advisor: Shlomo Zilberstein

Zoom link: https://umass-amherst.zoom.us/j/5266013533

Meeting ID: 526 601 3533

One tap mobile

+16468769923,,5266013533# US (New York)