engleski

ROBOT-ASSISTED AUTISM SPECTRUM DISORDER DIAGNOSTICS USING PARTIALLY OBSERVABLE MARKOV DECISION PROCESSES

The existing procedures for Autism Spectrum Disorder diagnosis are time consuming and challenging both for human evaluators and children being evaluated. The diagnosis of ASD relies solely on behavioral observations by experienced clinicians and occurrence of low agreement rates between different clinicians when evaluating a child suggests that there exists a need for a more objective approach to diagnostics and intervention. This thesis addresses that need by proposing a robot-assisted ASD diagnostic protocol consisting of four tasks adapted from ADOS. In this work the focus is on robot reasoning for ASD diagnostics. The main contribution of this thesis is a hierarchical Partially Observable Markov Decision Process framework that enables a humanoid robot to process the observations of child’s behavior, infer information about the unobservable state of the child and autonomously make decisions by selecting actions and tasks within the robot-assisted ASD diagnostic protocol. Each task of the protocol is modeled using a Mixed Observability Markov Decision Process model as a template. In order to formulate observation probabilities of task models, ASD experts are surveyed and their knowledge is encoded in the observation probabilities of task models. Expert knowledge also allowed for implementation of child behavioral models which are used to validate developed models. The model of the protocol is defined as a POMDP whose actions are tasks of the protocol. The interface between task and protocol models is formulated using regions of belief space of the task as observations for the protocol model. Following the successful validation through simulations with child behavioral models, task and protocol models are validated through experimental sessions with seven typically developing children and eight children with ASD. Results obtained through experiments show that the robot is capable of recognizing the behavior of the child and capable of differentiating different types of children, since the belief of the robot over the states of the child was comparable to assessment of autism experts.

robotics, autism spectrum disorder, diagnostics, partially observable Markov decision processes, mixed observability Markov decision processes, hierarchical framework, humanoid robot, autonomy

nije evidentirano