Reproduce “Language Understanding for Text-based Games Using Deep Reinforcement Learning” 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
Karthik Narasimhan, Tejas Kulkarni, Regina Barzilay
Researchers from MIT CSAIL developed an LSTM-DQN agent that learns to play text-based games by jointly inferring game state representations from raw textual descriptions and optimal action policies from game rewards. The model achieved 100% quest completion in a controlled environment and 96% in a complex Multi-User Dungeon, demonstrating robust language understanding without prior knowledge.
Key insights
- Learning expressive textual state representations directly from raw text using LSTMs is critical for effective language understanding and decision-making in text-based games.
- A deep neural network for the action scorer is essential for modeling complex control policies and value functions, surpassing simpler linear models.
- Learned textual representations are transferable across different game configurations, allowing agents to learn faster in new but similar environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.