Machine Learning Could Create the Perfect Game Bosses

game-aimachine-learningnpcgame-development

Abstraction: ML-driven NPC behavior in games balances difficulty tuning over scripted trees

Key points:

Eidos-Sherbrooke ML specialists demonstrated a Pac-Man-like game at Unreal Fest 2022 where enemy behavior is driven by a trained ML model rather than a behavior tree; Unreal Engine's EQS acts as the AI's sensory input
ML-optimal strategies (e.g., camping on collectibles) were found to be unenjoyable; the goal is not superhuman bots but tunable difficulty that remains fun
Traditional behavior trees require back-and-forth between designers and programmers; ML models give game designers a simpler dial for adjusting difficulty without diving into branching logic
Modl.ai uses ML bots to hunt graphical glitches, broken geometry, and unwinnable states during QA; prototyping cross-game deep learning to emulate real player behavior
Performance is a major barrier: large ML models inferring per game-clock tick are too slow; game engines (Unreal, Unity) are described as "anti-AI" in architecture
Physics-based animation is needed to handle unpredictable ML-driven movement, adding further hardware and development costs

Connections: Eidos Sherbrooke · Modl AI · Deepmind · Machine Learning · Game AI · Reinforcement Learning