Non-sequential Bayesian Multi-modal Inverse Reinforcement Learning

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2021-12-13

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

SCI3044

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

31+1

Series

Abstract

When modeling agents doing multiple or complex tasks, most research has done so in two stages, clustering the subtasks and then modeling each task independently of the others. Instead we propose a solution which would model both the tasks themselves and the dynamics of how the agent switches between these simultaneously. We experiment with multiple model architectures and progressively generalize the problem based on the number of tasks and switches between them. Our method successfully models a single reward function switch, but fails with multiple changes. We discuss why this happens and future directions of research to address these problems.

Description

Supervisor

Kaski, Samuel

Thesis advisor

Murena, Pierre-Alexandre

Keywords

inverse reinforcement learning, Bayesian, Markov chain Monte Carlo, multi-task IRL

Other note

Citation