Generic MCTS Planning for Versatile Agents

Abstract

The NPC Engine is a high-performance, domain-agnostic framework, written in Rust, for multi-agent simulations aimed at fostering emergent narrative. However, its traditional reliance on Rust programming and compile-time domain definitions presents challenges for rapid proto-typing and broader accessibility. This thesis addresses these limitations by investigating the suitability of Large Language Model (LLM) for generating simulation domains for the NPC Engine.

To enable this, a generalized interface was developed for the NPC Engine, facilitating runtime domain definition via JSON configuration and Lua scripting, while still demonstrating accept-able performance compared to Rust. The core contribution, the NPC Engine Lab, utilizes this interface and employs LLMs in a co-authoring process with built-in error correction to trans-late user prompts into complete simulation domains. A user study confirmed that LLMs are a promising tool for generating functional domains, effectively lowering entry barriers and accel-erating initial design, despite identified challenges in logical coherence.

This research validates a hybrid approach, combining LLM-based domain generation with Monte Carlo Tree Search (MCTS)-driven simulation, as a potent strategy for improving accessibility in complex simulation design.