David Ha • 2/15/2019

Learning Latent Dynamics for Planning from Pixels

This research paper presents the Deep Planning Network (PlaNet), a model-based agent that learns a latent dynamics model directly from pixel observations. It uses a combination of deterministic and stochastic transitions with a novel multi-step variational inference objective called latent overshooting. PlaNet solves complex continuous control tasks with sparse rewards and partial observability, achieving data efficiency far superior to model-free methods while matching or exceeding their final performance.

0 comments

#computer vision #Deep Learning #Reinforcement Learning