NPC engine: a High-Performance, Modular, and Domain-Agnostic MCTS Framework for Emergent Narrative

Abstract

The quality of emergent narrative is strongly linked to the capabilities of the underlying simulation. A powerful way to drive character activity is to use a performant and dependable planner. In this paper, we present a domain-agnostic multi-agent Monte Carlo tree search planner implemented using the Rust programming language. The planner supports tasks of varying duration and a dynamic number of agents with heterogeneous properties. It also provides optional concurrency, allowing for scalable simulations of many agents planning in parallel on multiple threads. In addition to simulation-based rollout, it supports custom state value estimators and offers a basic adaptive implementation using neural networks. The planner also includes a variety of debugging features, such as the ability to plot the search tree. For easy adoption, we provide several documented application examples.