Adversarial Autoencoder and Multi-Armed Bandit for Dynamic Difficulty Adjustment in Immersive Virtual Reality for...

Area of Science:

• * Neuroscience
• * Machine Learning
• * Rehabilitation Engineering

Background:

• * Motor rehabilitation requires frequent, time-consuming adjustments based on patient progress.
• * High-dimensional data from motor control tasks presents analysis challenges.
• * Virtual reality (VR) offers immersive environments for motor skill assessment.

Purpose of the Study:

• * To develop an efficient tool for dimensionality reduction of hand movement data in VR.
• * To create a visualization tool using a latent space representation.
• * To integrate this tool into a reinforcement learning framework for decision-making in rehabilitation.

Main Methods:

• * Hand movement data was collected using a wireless controller with Oculus Rift S in a VR environment.
• * Adversarial Autoencoder (AAE) models (unsupervised and semi-supervised) were used for dimensionality reduction.
• * Multi-Armed Bandit (MAB) algorithms (Boltzmann and Sibling Kalman filters) utilized latent space distance as rewards.

Main Results:

• * Adversarial Autoencoder models successfully reduced high-dimensional hand movement data into a 2D latent space.
• * Multi-Armed Bandit agents demonstrated efficient learning of distance evolution within the latent space.
• * Sibling Kalman filter exploration significantly outperformed Boltzmann exploration in accuracy.

Conclusions:

• * The proposed deep learning approach effectively visualizes and tracks motor control capabilities in a VR context.
• * This tool can aid clinicians in making timely therapy adjustments, improving patient outcomes.
• * The integration of AAE and MAB offers a promising framework for data-driven motor rehabilitation.