У нас вы можете посмотреть бесплатно Rich Sutton, Planning and Action Selection in Options-based Agents или скачать в максимальном доступном качестве, видео которое было загружено на ютуб. Для загрузки выберите вариант из формы ниже:
Если кнопки скачивания не
загрузились
НАЖМИТЕ ЗДЕСЬ или обновите страницу
Если возникают проблемы со скачиванием видео, пожалуйста напишите в поддержку по адресу внизу
страницы.
Спасибо за использование сервиса ClipSaver.ru
Rich Sutton is here to challenge one of the biggest misconceptions in reinforcement learning. In his talk, "Planning and Action Selections in Options-based Agents," the Turing Award winner makes a bold claim: when designing AI, you should never "execute" an option to completion. You should only activate it. Why? Because the real world is unpredictable. Surprising events requires the flexibility to abandon a plan and react instantly. While options are invaluable for representing knowledge and planning at a high level, Sutton argues that committing to them is a mistake. Sutton delves into the core challenges of temporal abstraction, the limits of traditional planning methods, and the deep theoretical connections between planning and learning. He presents a new way of thinking about how agents should behave in a complex, dynamic world. Timestamps: 0:00 - Introduction and a bold claim 3:05 - Options as a mechanism for temporal abstraction 8:47 - The Common Model of the Intelligent Agent 15:56 - The agent architecture with multiple policies and value functions 23:50 - The generality of all planning methods 30:14 - The crucial difference between activation and execution 37:59 - The Spy Plane example: a practical application 44:04 - Exploration and intrinsic reward 46:52 - Deeply hierarchical policies and the option keyboard 49:37 - Conclusion 50:00 - Q&A #RichSutton #ReinforcementLearning #ArtificialIntelligence #AI #MachineLearning #TemporalAbstraction #OptionsFramework #SuttonBarto #DeepLearning #Research