Interactive Decision Making, Influence-based Abstraction & Transfer Planning

13^th October 2015, 13:00 Ashton Lecture Theater
Dr Frans Oliehoek
Department of Computer Science
Liverpool University

Abstract

In this talk, I will take the opportunity to briefly introduce myself and give a quick overview of some different problems in the field of 'interactive decision making' that I have worked on, such as: planning for robotic task allocation problems, seller selection in e-commerce settings, and learning about other agents in partially observable settings.

Subsequently I will dive a bit deeper in one of the main topics that I have investigated over the last few years: planning for large factored decentralized partially observable Markov decision processes (factored Dec-POMDPs). These models provide a generic and principled framework for multiagent decision making problems under uncertainty, but are notoriously difficult to solve. I will introduce a heuristic approach, called transfer planning (TP), talk about a way of doing abstraction in these problems and how this leads to informative upper bounds on the quality of TP.