Title:	Towards the fast inference of Diffusion Models for Games
DNr:	Berzelius-2025-103
Project Type:	LiU Berzelius
Principal Investigator:	Ruoqi Zhang <ruoqi.zhang@it.uu.se>
Affiliation:	Uppsala universitet
Duration:	2025-03-12 – 2025-10-01
Classification:	10210
Keywords:

Abstract

This research project investigates advanced reinforcement learning techniques for gaming environments, specifically focusing on enhancing consistency policies for offline reinforcement learning. Modern game AI requires both expressive action modeling and fast inference - requirements that are often at odds with each other. While diffusion models offer high expressivity through multi-step denoising processes, they impose significant computational burdens that limit real-time application. Consistency policies provide an attractive alternative by enabling single-step inference, substantially improving sampling speed, but often demonstrate lower expressiveness when trained via pure imitation learning. The main idea now is to addresses this fundamental tradeoff by developing enhanced consistency policies that integrate Q-learning and ensemble techniques.