Reproduce “Baselines for Reinforcement Learning in Text Games” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mikuláš Zelinka
A new deep reinforcement learning architecture, the Siamese State-Action Q-Network (SSAQN), was developed for playing text-based games, achieving optimal performance on deterministic games and improved results on non-deterministic ones compared to baselines. This work also released `pyfiction`, an open-source library that standardizes access to various interactive fiction environments for research.
Key insights
- A siamese architecture with shared embedding and LSTM weights can effectively learn joint representations for natural language states and actions in text games.
- Training an agent on multiple diverse text games simultaneously can lead to improved performance on non-deterministic games, suggesting beneficial cross-game knowledge transfer.
- Broad generalization to entirely unseen text games is challenging with limited datasets, indicating a need for more extensive lexical and semantic overlap between training and test environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.