Curiosity-driven Exploration by Self-supervised Prediction
Reinforcement Learning / Curiosity

Curiosity-driven Exploration by Self-supervised Prediction

Abstract

在现实世界中自监督预测的好奇心驱动的探索，能够提供给agent的奖励非常少，或者根本不存在。在这种情况下，好奇心可以作为一种内在的奖励信号，使agent能够探索环境并学习可能在以后的生活中有用的技能。本文把好奇心表述为一个agent在视觉特征空间预测自身行为结果的能力时的误差，并把这种error作为内在好奇心reward提供给agent进行学习，该视觉特征空间是由一个自监督逆动态模型学习的。

Read More
Follow

Links

Categories

Tag Cloud

A3C Actor-Critic Curiosity Curriculum Learning Data Augmentation Dialogue system Latent Variable Memory Network Multi-task Learning NLU Paper overview Reinforcement Learning Self-paced Learning User Simulator Visual Dialog jupyter meta-learning reward word2vec

Recent

Archives

Tags

A3C 1
Actor-Critic 1
Curiosity 1
Curriculum Learning 3
Data Augmentation 1
Dialogue system 2
Latent Variable 1
Memory Network 1
Multi-task Learning 2
NLU 1
Paper overview 1
Reinforcement Learning 6
Self-paced Learning 1
User Simulator 1
Visual Dialog 1
jupyter 1
meta-learning 1
reward 1
word2vec 1

Tag Cloud

A3C Actor-Critic Curiosity Curriculum Learning Data Augmentation Dialogue system Latent Variable Memory Network Multi-task Learning NLU Paper overview Reinforcement Learning Self-paced Learning User Simulator Visual Dialog jupyter meta-learning reward word2vec

Recent

Archives

Tags

A3C 1
Actor-Critic 1
Curiosity 1
Curriculum Learning 3
Data Augmentation 1
Dialogue system 2
Latent Variable 1
Memory Network 1
Multi-task Learning 2
NLU 1
Paper overview 1
Reinforcement Learning 6
Self-paced Learning 1
User Simulator 1
Visual Dialog 1
jupyter 1
meta-learning 1
reward 1
word2vec 1
+
Your browser is out-of-date!

Update your browser to view this website correctly. Update my browser now

×