Скачать с ютуб World-Models 🌍 Model Based Reinforcement Learning в хорошем качестве

World-Models 🌍 Model Based Reinforcement Learning

In this video we talk about architectures for learning an accurate model of the environment, that is then be used by the agent for learning a policy or a value function, that is, for doing reinforcement learning. We discuss two important factors that play an important role in the correct choice of the architecture in model based reinforcement learning: stochasticity and partial observability. The former is solved by state space models, the latter by autoregressive models. The video provide a in-depth review of these models with a particular emphasis on their use in model based reinforcement learning. Enjoy the video! Resources: Dream To Control: Learning Behaviors By Latent Imagination https://arxiv.org/pdf/1912.01603.pdf Learning and Querying Fast Generative Models for Reinforcement Learning https://arxiv.org/pdf/1802.03006.pdf Stochastic Variational Video Prediction https://arxiv.org/pdf/1710.11252.pdf 0:00 Introduction 2:44 What's a World Model? 4:26 Dealing with partial observability 5:44 Autoregressive Models 6:39 Recurrent Autoregressive Models 8:06 Environments are Stochastic 11:02 Stochastic State Space Models 14:00 Recurrent Stochastic State Space Models 14:47 Recurrent State Space Models 18:00 Discussion #WorldModels, #ReinforcementLearning, #StateSpaceModels, #RNN, #VariationalMethods, #Model-BasedRL, #BayesianDeepLearning #ModelBasedReinforcementLearning