From Player to System: An Agent-Based Framework for Modeling Human Performance in Tennis

Keywords: Agent-based modeling, Human performance, Biomechanics, Cognitive science, Sports science, Skill acquisition, AI for science, Reinforcement learning

TL;DR: Modeling athlete transformation as an agent-based system that unifies biomechanics, cognition, and psychology into a reproducible training loop for human performance science.

Abstract:

We present an agent-based framework for studying human performance as an interacting multi-agent system, instantiated on tennis skill acquisition from novice (NTRP 2.0) to advanced amateur (5.0). We formalize the athlete as an adaptive agent optimizing a reward that composes biomechanical efficiency, perceptual-cognitive anticipation, and psychological regulation, under constraints of injury risk and workload. The environment comprises task contexts (serve, return, baseline exchanges), surfaces, loads, and opponent models. We operationalize state with multi-modal signals (kinematics, kinetics, gaze/oculomotor proxies, and psychometrics), actions as motor programs and pre-shot routines, and rewards as performance, energy economy, and safety proxies.

Building on literature in kinetic-chain biomechanics, perceptual-cognitive expertise, and periodized training, we propose an Agentic Training Loop: (1) sense and attribute bottlenecks, (2) generate intervention hypotheses, (3) schedule and dose training, and (4) evaluate with task-level and physiology-level metrics. We release an open protocol (schemas, prompts, and evaluation harness) and a gym-style simulation for in-silico ablation before human trials. The framework is domain-general and extends to rehabilitation and broader human skill science. This is an AI-led study: an AI scientist agent generated hypotheses, designed the methodology, synthesized literature, and drafted the manuscript; human co-authors provided domain constraints and ethical oversight. Our results are a reproducible blueprint for agentic science of human performance.

Submission Number: 169