Adaptive Prompt Selection and Fading Optimization for Autism Skill Acquisition: A Reinforcement Learning Approach

Abstract

This research addresses critical challenges in autism spectrum disorder skill instruction through computational optimization of prompting strategies. Discrete trial training relies heavily on systematically prompt delivery and fading, yet practitioners lack algorithmic guidance for optimal decision-making. We formalize prompt selection as a Markov decision process and develop three adaptive algorithms: threshold-based rules, progressive delay optimization, and Q-learning variants. Simulated teaching data from 120 learners with varying response profiles validate algorithm performance across skill acquisition speed, error patterns, and independence promotion. Results demonstrate that Q-learning with ε-greedy exploration reduces trials to mastery by 23.7% compared to fixed-schedule baselines, while achieving 84.6% generalization accuracy and 91.2% one-week maintenance accuracy. Learner-specific feature matching achieves 87.4% prediction accuracy for optimal strategy selection. These findings provide evidence-based algorithmic frameworks for applied behavior analysis practitioners to enhance instructional efficiency while minimizing prompt dependency risks.