microsoft/jericho

What it solves

Jericho has created a python-based interface that allows learning agents to interact with interactive fiction games. This provides a benchmark for AI agents to practice reading comprehension and strategic decision-making within text-based environments.

How it works

The library provides a FrotzEnv environment, which acts as a wrapper around the emulator emulator for interactive fiction games. It allows agents to send text commands and receive text observations. It includes tools for managing game seeds to ensure experiments are reproducible or stochastic (randomized), and provides a template action generator to help agents formulate valid game commands.

Who it’s for

Researchers and developers working on reinforcement learning (RL) and natural language processing (NLP) who want to test their agents' ability to solve complex text-based games.

Highlights

• Reproducible Experiments: Supports explicit seeding and walkthrough seeds to reproduce specific game states.
• Pythonic Interface: Lightweight Python interface for connecting agents to games.
• Integration with RL Agents: Compatible with various research agents like RCDQN, CALM, and KG-A2C.
• Game State Management: Includes an object tree and game dictionary for better environment interaction.