Graph Networks as Learnable Physics Engines for Inference and Control

Alvaro Sanchez-Gonzalez, Nicolas Heess, Jost Tobias Springenberg, Josh Merel, Martin Riedmiller, Raia Hadsell, Peter Battaglia
Proceedings of the 35th International Conference on Machine Learning, PMLR 80:4470-4479, 2018.

Abstract

Understanding and interacting with everyday physical scenes requires rich knowledge about the structure of the world, represented either implicitly in a value or policy function, or explicitly in a transition model. Here we introduce a new class of learnable models–based on graph networks–which implement an inductive bias for object- and relation-centric representations of complex, dynamical systems. Our results show that as a forward model, our approach supports accurate predictions from real and simulated data, and surprisingly strong and efficient generalization, across eight distinct physical systems which we varied parametrically and structurally. We also found that our inference model can perform system identification. Our models are also differentiable, and support online planning via gradient-based trajectory optimization, as well as offline policy optimization. Our framework offers new opportunities for harnessing and exploiting rich knowledge about the world, and takes a key step toward building machines with more human-like representations of the world.

Cite this Paper


BibTeX
@InProceedings{pmlr-v80-sanchez-gonzalez18a, title = {Graph Networks as Learnable Physics Engines for Inference and Control}, author = {Sanchez-Gonzalez, Alvaro and Heess, Nicolas and Springenberg, Jost Tobias and Merel, Josh and Riedmiller, Martin and Hadsell, Raia and Battaglia, Peter}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {4470--4479}, year = {2018}, editor = {Dy, Jennifer and Krause, Andreas}, volume = {80}, series = {Proceedings of Machine Learning Research}, month = {10--15 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v80/sanchez-gonzalez18a/sanchez-gonzalez18a.pdf}, url = {https://proceedings.mlr.press/v80/sanchez-gonzalez18a.html}, abstract = {Understanding and interacting with everyday physical scenes requires rich knowledge about the structure of the world, represented either implicitly in a value or policy function, or explicitly in a transition model. Here we introduce a new class of learnable models–based on graph networks–which implement an inductive bias for object- and relation-centric representations of complex, dynamical systems. Our results show that as a forward model, our approach supports accurate predictions from real and simulated data, and surprisingly strong and efficient generalization, across eight distinct physical systems which we varied parametrically and structurally. We also found that our inference model can perform system identification. Our models are also differentiable, and support online planning via gradient-based trajectory optimization, as well as offline policy optimization. Our framework offers new opportunities for harnessing and exploiting rich knowledge about the world, and takes a key step toward building machines with more human-like representations of the world.} }
Endnote
%0 Conference Paper %T Graph Networks as Learnable Physics Engines for Inference and Control %A Alvaro Sanchez-Gonzalez %A Nicolas Heess %A Jost Tobias Springenberg %A Josh Merel %A Martin Riedmiller %A Raia Hadsell %A Peter Battaglia %B Proceedings of the 35th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2018 %E Jennifer Dy %E Andreas Krause %F pmlr-v80-sanchez-gonzalez18a %I PMLR %P 4470--4479 %U https://proceedings.mlr.press/v80/sanchez-gonzalez18a.html %V 80 %X Understanding and interacting with everyday physical scenes requires rich knowledge about the structure of the world, represented either implicitly in a value or policy function, or explicitly in a transition model. Here we introduce a new class of learnable models–based on graph networks–which implement an inductive bias for object- and relation-centric representations of complex, dynamical systems. Our results show that as a forward model, our approach supports accurate predictions from real and simulated data, and surprisingly strong and efficient generalization, across eight distinct physical systems which we varied parametrically and structurally. We also found that our inference model can perform system identification. Our models are also differentiable, and support online planning via gradient-based trajectory optimization, as well as offline policy optimization. Our framework offers new opportunities for harnessing and exploiting rich knowledge about the world, and takes a key step toward building machines with more human-like representations of the world.
APA
Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R. & Battaglia, P.. (2018). Graph Networks as Learnable Physics Engines for Inference and Control. Proceedings of the 35th International Conference on Machine Learning, in Proceedings of Machine Learning Research 80:4470-4479 Available from https://proceedings.mlr.press/v80/sanchez-gonzalez18a.html.

Related Material