Graph networks as learnable physics engines for inference and control

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.

arXiv commentICML 2018

References in corpus (22)

• Auto-Encoding Variational Bayes
• Continuous control with deep reinforcement learning
• Semi-Supervised Classification with Graph Convolutional Networks
• Neural Message Passing for Quantum Chemistry
• Geometric deep learning: going beyond Euclidean data
• Spectral Networks and Locally Connected Networks on Graphs
• Stochastic Backpropagation and Approximate Inference in Deep Generative Models
• Deep Convolutional Networks on Graph-Structured Data
• On the Properties of Neural Machine Translation: Encoder-Decoder Approaches
• Learning Convolutional Neural Networks for Graphs
• DeepMind Control Suite
• A simple neural network module for relational reasoning
• Discriminative Embeddings of Latent Variable Models for Structured Data
• Gated Graph Sequence Neural Networks
• Learning Continuous Control Policies by Stochastic Value Gradients
• EPOpt: Learning Robust Neural Network Policies Using Model Ensembles
• Learning Visual Predictive Models of Physics for Playing Billiards
• Learning model-based planning from scratch
• Discovering objects and their relations from entangled scene representations
• Metacontrol for Adaptive Imagination-Based Optimization
• Planning to Be Surprised: Optimal Bayesian Exploration in Dynamic Environments
• Learning Awareness Models

Cited by in corpus (73)

• Learning to Simulate Complex Physics with Graph Networks
• Contrastive Multi-View Representation Learning on Graphs
• Learning to schedule job-shop problems: Representation and policy learning using graph neural network and reinforcement learning
• Discovering Symbolic Models from Deep Learning with Inductive Biases
• Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids
• Universal Invariant and Equivariant Graph Neural Networks
• Algorithmic Framework for Model-based Deep Reinforcement Learning with Theoretical Guarantees
• GAG: Global Attributed Graph Neural Network for Streaming Session-based Recommendation
• Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning
• Contrastive Learning of Structured World Models
• Graph Neural Ordinary Differential Equations
• Port-Hamiltonian Neural Networks for Learning Explicit Time-Dependent Dynamical Systems
• Learning Compositional Koopman Operators for Model-Based Control
• Differentiable Physics-informed Graph Networks
• Causal Discovery in Physical Systems from Videos
• Self-supervised Auxiliary Learning with Meta-paths for Heterogeneous Graphs
• Recurrent Event Network: Autoregressive Structure Inference over Temporal Knowledge Graphs
• Finite Volume Graph Network(FVGN): Predicting unsteady incompressible fluid dynamics with finite volume informed neural network
• Learning Symbolic Physics with Graph Networks
• One Policy to Control Them All: Shared Modular Policies for Agent-Agnostic Control
• Context-aware Dynamics Model for Generalization in Model-Based Reinforcement Learning
• Graph Neural Networks for Leveraging Industrial Equipment Structure: An application to Remaining Useful Life Estimation
• Streaming Graph Neural Networks
• Neural Information Squeezer for Causal Emergence
• Visual Grounding of Learned Physical Models
• Graph-Structured Visual Imitation
• Approximating the solution to wave propagation using deep neural networks
• Multi-Stage Self-Supervised Learning for Graph Convolutional Networks on Graphs with Few Labels
• ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics
• Reliable Graph Neural Networks for Drug Discovery Under Distributional Shift
• AliGraph: A Comprehensive Graph Neural Network Platform
• 3D-OES: Viewpoint-Invariant Object-Factorized Environment Simulators
• A PAC-Bayesian Approach to Generalization Bounds for Graph Neural Networks
• Active Learning of Abstract Plan Feasibility
• Learning to simulate and design for structural engineering
• Scalable variational Monte Carlo with graph neural ansatz
• Improving Generative Imagination in Object-Centric World Models
• Graph Convolutional Networks with EigenPooling
• Constant Time Graph Neural Networks
• Learning continuous-time PDEs from sparse data with graph neural networks
• Are Graph Neural Networks Miscalibrated?
• Trajectory-wise Multiple Choice Learning for Dynamics Generalization in Reinforcement Learning
• Learning Self-Correctable Policies and Value Functions from Demonstrations with Negative Sampling
• Physics Encoded Blocks in Residual Neural Network Architectures for Digital Twin Models
• Modeling Attention Flow on Graphs
• Comparing Semi-Parametric Model Learning Algorithms for Dynamic Model Estimation in Robotics
• PhysGNN: A Physics-Driven Graph Neural Network Based Model for Predicting Soft Tissue Deformation in Image-Guided Neurosurgery
• Task-Agnostic Morphology Evolution
• Graph Neural Processes: Towards Bayesian Graph Neural Networks
• Accurately Solving Physical Systems with Graph Learning
• Towards Scale-Invariant Graph-related Problem Solving by Iterative Homogeneous Graph Neural Networks
• A Model-Constrained Tangent Slope Learning Approach for Dynamical Systems
• Continuous-Depth Neural Models for Dynamic Graph Prediction
• Physics-aware Spatiotemporal Modules with Auxiliary Tasks for Meta-Learning
• Isometric Transformation Invariant and Equivariant Graph Convolutional Networks
• Causal World Models by Unsupervised Deconfounding of Physical Dynamics
• Hybrid modeling: Applications in real-time diagnosis
• Explicit Pairwise Factorized Graph Neural Network for Semi-Supervised Node Classification
• HR-RCNN: Hierarchical Relational Reasoning for Object Detection
• Graph Convolutional Networks in Feature Space for Image Deblurring and Super-resolution
• Interpretable Drug Synergy Prediction with Graph Neural Networks for Human-AI Collaboration in Healthcare
• Evaluating the Apperception Engine
• Modular meta-learning in abstract graph networks for combinatorial generalization
• Neural Consciousness Flow
• Model discovery in the sparse sampling regime
• Bayes EMbedding (BEM): Refining Representation by Integrating Knowledge Graphs and Behavior-specific Networks
• Fractional order graph neural network
• Convolutions for Spatial Interaction Modeling
• A Neural-Symbolic Framework for Mental Simulation
• Biological Blueprints for Next Generation AI Systems
• A Multi-Semantic Metapath Model for Large Scale Heterogeneous Network Representation Learning
• Cross-lingual Transfer for Text Classification with Dictionary-based Heterogeneous Graph
• Can NetGAN be improved on short random walks?