Reinforcement Learning on Web Interfaces Using Workflow-Guided Exploration

Evan Zheran Liu , Kelvin Guu , Panupong Pasupat , Tianlin Shi , Percy Liang

🏛 Institutions: Stanford
📅 Date: February 24, 2018
📑 Publisher: ICLR 2018 (Poster)
💻 Env: Web
🔑 Keywords: framework reinforcement learning workflow-guided exploration MiniWoB

TLDR

This paper introduces workflow-guided exploration, where demonstrations are converted into high-level workflows that constrain web-agent exploration during RL. It shows that this makes sparse-reward web interaction much more sample-efficient than pure behavioral cloning on World of Bits style tasks.

Open paper Report issue