This website requires JavaScript.

Guaranteed Discovery of Controllable Latent States with Multi-Step Inverse Models

Alex LambRiashat IslamYonathan Efroni ...+6 John Langford
Jul 2022
摘要
A person walking along a city street who tries to model all aspects of theworld would quickly be overwhelmed by a multitude of shops, cars, and peoplemoving in and out of view, following their own complex and inscrutabledynamics. Exploration and navigation in such an environment is an everydaytask, requiring no vast exertion of mental resources. Is it possible to turnthis fire hose of sensory information into a minimal latent state which isnecessary and sufficient for an agent to successfully act in the world? Weformulate this question concretely, and propose the Agent-Controllable StateDiscovery algorithm (AC-State), which has theoretical guarantees and ispractically demonstrated to discover the \textit{minimal controllable latentstate} which contains all of the information necessary for controlling theagent, while fully discarding all irrelevant information. This algorithmconsists of a multi-step inverse model (predicting actions from distantobservations) with an information bottleneck. AC-State enables localization,exploration, and navigation without reward or demonstrations. We demonstratethe discovery of controllable latent state in three domains: localizing a robotarm with distractions (e.g., changing lighting conditions and background),exploring in a maze alongside other agents, and navigating in the Matterporthouse simulator.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答