Reproduce “Playing Text-Adventure Games with Graph-Based 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

Prithviraj Ammanabrolu, Mark O. Riedl

Researchers at the Georgia Institute of Technology developed Knowledge Graph DQN (KG-DQN), an architecture for playing text-adventure games that utilizes a knowledge graph for state representation and action pruning, alongside transfer learning from a question-answering task. This system demonstrated faster reward convergence and comparable quest completion efficiency on TextWorld games when contrasted with existing deep reinforcement learning baselines.

Key insights

  • Explicitly structured knowledge graphs serve as an effective persistent memory for RL agents, enabling them to maintain long-term context and reason about relationships in partially observable text environments.
  • The combination of knowledge graph-driven action pruning and pre-training via question answering improves both the efficiency of learning (faster reward convergence) and the quality of quest solutions (fewer steps to completion).
  • Framing game play as a question-answering task allows for valuable knowledge transfer from other games, accelerating policy learning in new, unseen environments without requiring specific prior quest knowledge.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.