A Rapid and Automated Difficulty Evaluation Framework for Rhythm Games via Persona-Driven Multi-Task Behavioral Cloning

초록

Rhythm games require frequent beatmap difficulty balancing, but manual quality assurance is time-consuming and subjective, while heuristic metrics such as notes per second fail to capture cognitive and biomechanical load. We propose an automated difficulty evaluation framework driven by a Multi-Task Behavioral Cloning Network (MT-BCN) that parameterizes three player personas: beginner, normal, and expert. The beginner agent quantifies cognitive overload via miss rates in dense patterns, while the normal and expert agents measure biomechanical execution difficulty through timing offsets. Persona-derived terms are fused into a complexity score and quantized into a fifteen-tier hierarchy. On forty-five beatmaps, the framework achieves a Spearman rank correlation of ρ =0.949 against offline proxy tiers in approximately 4.5 seconds end-to-end. To address the inherent circularity of proxy-based validation, a two-stage human study showed low inter-rater reliability under chart-based Likert scoring (ICC(2,k)=0.227 , five annotators) but 74.1% agreement in a nine-rater pairwise verification of the extreme tiers (p< 10-5). As a circularity-free check, we additionally applied the trained framework zero-shot to 287 charts from two independent commercial games (osu!mania and StepMania), whose difficulty labels are assigned independently of our pipeline; the hybrid ranking attained Spearman ρ =0.88 -0.98 across both 4-key and 7-key layouts, with mid-tier pairwise agreement of 95-100%. The pipeline substantially reduces iterative evaluation time within our internal workflow, providing a complementary evaluation aid; final balancing and player-experience review remain a human responsibility. © 2026 The Authors.

키워드