Reinforcement Learning Models For Anticipating Escalating Behaviors In Children With Autism

Abstract

Autism Spectrum Disorder (ASD) is a challenging condition for which it is not possible to predict and manage behavioural escalations, including aggressiveness, self-injury or outbursts. Traditional observational approaches tend to be not scalable and responsive in real-time. This paper describes the use of reinforcement learning (RL) agents in predicting the escalating behaviors of children with autism, using wearable IoT sensors and data about the context. A multi-agent RL is employed where continuous sensor feedback is available for the prediction model. The approach is a combination of Q-learning and deep reinforcement learning (DRL) which lets us predict thermodynamic trends towards a breakdown and inspires a rapid alert function. It is shown that the proposed system has better predictability than the baseline machine learning classifiers. This research represents how RL-based adaptive models can support caregivers and clinicians for proactive intervention planning while the care plan limits caregivers' and families' stress for children.