NetHack Strategy API

Overview

This repository builds upon the NetPlay repository by extending its capabilities with strategy abstraction for guiding agents through high-level gameplay objectives. The Strategy API enhances the existing framework by introducing reusable strategies to enable more effective exploration of NetHack's early game. Also building out a Skill API that is a wrapper for the RL environment, not fully completed.

---

Available Strategies

The following early game strategies are available for guiding the agent:

• level_5: Reach experience level 5+ without descending past Dungeon Level 5.
• stash: Create a stash near a safe location (e.g., stairs or an altar).
• clear_minetown: Clear Minetown and benefit from its resources.
• clear_sobokan: Complete the Sokoban puzzle branch to obtain rewards.
• find_use_altars: Find and use an altar for identifying items and making sacrifices.
• clear_gnomish_mines: Safely clear the Gnomish Mines and retrieve key items.

Each strategy includes specific tasks and tips designed to help agents succeed in early game exploration and survival.

---

Setup

Follow setup requirements in the following repositories:

• NetPlay
• NLE
• nle-language-wrapper

---

Usage

Examples: Run the LLM agent from NetPlay with default settings:

python ./run_full_runs.py llm -agent_name "NetPlay_agent" -num_runs 1 -model "gpt-4o-mini" --use_guide --render

Run the LLM agent with stash strategy (or any other) AND general prompt:

python ./run_strategies.py llm -agent_name "stash" -strategy "stash" -num_runs 1 -model "gpt-4o-mini" --use_guide --render

Run the LLM agent with stash strategy (or any other) ONLY (no general prompt):

python ./run_strategies.py llm -agent_name "stash" -strategy "stash" -num_runs 1 -model "gpt-4o-mini" --render

---

Citations

1. Nikolaj Goodger, Peter Vamplew, Cameron Foale, and Richard Dazeley.
   A NetHack learning environment language wrapper for autonomous agents.
   Journal of Open Research Software, 11, June 2023.

2. Dominik Jeurissen, Diego Perez-Liebana, Jeremy Gow, Duygu Cakmak, and James Kwan.
   Playing NetHack with LLMs: Potential & Limitations as Zero-Shot Agents.
   In 2024 IEEE Conference on Games (CoG), pages 1–8, 2024.

3. Heinrich Küttler, Nantas Nardelli, Alexander H. Miller, Roberta Raileanu, Marco Selvatici, Edward Grefenstette, and Tim Rocktäschel.
   The NetHack Learning Environment.
   In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2020.