Sum–product networks for learning from demonstration with tractable exact inference and Bayesian experimental design
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Science |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Mcode
Language
en
Pages
46
Series
Abstract
One of the main challenges for robotic systems is to operate reliably in physical environments with latent (unobserved) parameters. Learning from Demonstration (LfD) offers a practical approach to tackle this problem by acquiring policies from expert demonstrations; however, limited data makes learned policies sensitive to data scarcity and distribution shifts. Bayesian Experimental Design (BED) addresses this by actively selecting queries that are maximally informative about latent environment parameters, which typically requires many repeated evaluations of conditional and marginal distributions. This thesis investigates whether Sum-Product Networks (SPNs) can serve as a tractable generative model that supports tractable likelihood evaluation, efficient conditioning, and marginalization over arbitrary subsets of trajectory variables. We study this question through two main proofs of concept: a) Synthetic Van der Pol and Lotka–Volterra systems, and real handwriting trajectories from the LASA dataset; b) a toy Bayesian Experimental Design for Learning from Demonstration tasks, demonstrating that SPNs’ exact inference enables accurate and reliable EIG calculations, leading to more efficient query selection. Across these experiments, the SPN model in this thesis matches the key qualitative structure of the studied nonlinear systems, supports inpainting of missing segments under the considered conditioning patterns, and enables numerically stable EIG computation via exact likelihood and marginal queries for the studied BED settings. Overall, the results provide initial evidence that SPNs can support tractable trajectory-based inference and small-scale BED in Lfd, while scaling to higher-dimensional observations and design spaces remains future work.Description
Supervisor
Kyrki, VilleThesis advisor
Keurulainen, OskarShen, Zheng