AIGA: AI‑Generated Action for Game AI Bots – From AIGC to Human‑like and Stylized Agents

This article presents the work of NetEase Fuxi on AI‑Generated Action (AIGA), a form of AIGC that generates agent actions for game AI bots.

It outlines three main topics: the transition from AIGC to AIGA, research progress on human‑like (personification) and stylized game bots, and the use of RLHF to fine‑tune bot models.

The authors describe the importance of personification and style for player flow, propose quantitative metrics (objective, diversity, competitiveness, domain‑specific), and show how radar‑chart visualisation can compare algorithms.

They discuss combining imitation learning with reinforcement learning, using reward shaping and automated weight adjustment to balance strength and human‑likeness, and present case studies in games such as FIFA, Diablo, Elden Ring, and basketball titles.

Experiments on multi‑style bots (e.g., “Active Bot” vs “Lazy Bot”) and RLHF‑based fine‑tuning demonstrate improvements in objective metrics but also highlight gaps between objective scores and subjective player perception.

The paper concludes with a pipeline that integrates data collection, pre‑training, self‑imitation, RLHF, and iterative human feedback to create more generalizable, automatically optimized game AI agents.