Robust Locally-Linear Controllable Embedding

Abstract

Embed-to-control (E2C) is a model for solving high-dimensional optimal control problems by combining variational auto-encoders with locally-optimal controllers. However, the E2C model suffers from two major drawbacks: 1) its objective function does not correspond to the likelihood of the data sequence and 2) the variational encoder used for embedding typically has large variational approximation error, especially when there is noise in the system dynamics. In this paper, we present a new model for learning robust locally-linear controllable embedding (RCE). Our model directly estimates the predictive conditional density of the future observation given the current one, while introducing the bottleneck between the current and future observations. Although the bottleneck provides a natural embedding candidate for control, our RCE model introduces additional specific structures in the generative graphical model so that the model dynamics can be robustly linearized. We also propose a principled variational approximation of the embedding posterior that takes the future observation into account, and thus, makes the variational approximation more robust against the noise. Experimental results show that RCE outperforms the E2C model, and does so significantly when the underlying dynamics is noisy.

13 pages

Cited by in corpus (15)

• Dream to Control: Learning Behaviors by Latent Imagination
• Learning Robotic Manipulation through Visual Planning and Acting
• Offline Reinforcement Learning from Images with Latent Space Models
• Adaptive Path-Integral Autoencoder: Representation Learning and Planning for Dynamical Systems
• Learning Semantic Embedding Spaces for Slicing Vegetables
• The Differentiable Cross-Entropy Method
• Robot Motion Planning in Learned Latent Spaces
• Models, Pixels, and Rewards: Evaluating Design Trade-offs in Visual Model-Based Reinforcement Learning
• Video Extrapolation with an Invertible Linear Embedding
• Planning in Learned Latent Action Spaces for Generalizable Legged Locomotion
• Learning Latent State Spaces for Planning through Reward Prediction
• Disentangling Dynamics and Content for Control and Planning
• Market Self-Learning of Signals, Impact and Optimal Trading: Invisible Hand Inference with Free Energy
• Learn Proportional Derivative Controllable Latent Space from Pixels
• End-to-end neural network approach to 3D reservoir simulation and adaptation