Adaptive Self-Efficacy-Based AI Coaching for Cycling

The Thompson Sampling contextual bandit algorithm, trained on Session 1 data, monitors performance continuously and evaluates every 5 seconds whether to deliver an affirmation. The policy is trained to maximize a multi-objective "efficacy-preserving performance" function that rewards:

• Maintaining target power relative to rolling 30s/2min/5min baselines
• Stabilizing short-horizon power variability (30s coefficient of variation)
• Stabilizing heart-rate (HR) trajectory consistent with efficient pacing

The decision process considers:

• Current power relative to 30-second, 2-minute, and 5-minute rolling averages
• Power output variability (coefficient of variation over past 30 seconds)
• Heart rate trajectory and cardiac drift patterns
• Cadence stability and changes from baseline
• Time elapsed and expected fatigue progression based on power-duration curve Self-efficacy-based AI coaching adapts to physiological measures (power and heart rate).

Generic motivational messages delivered at fixed intervals (minutes 3, 6, 9, 12, 15, and 18) regardless of performance state. Messages follow the same complexity gradient based on elapsed time rather than individual response:

• Minutes 3, 6: "You're building momentum with every pedal stroke-maintain this strong rhythm"
• Minutes 9, 12: "Strong effort-push through this challenge"
• Minutes 15, 18: "Final push-finish strong"