Motor Learning and Machine Learning: Predicting the Amount of Sessions to Learn the Tracing Task

Abstract

Motor activities are the main way of interacting with the world. Therefore, loss of ability to perform some of these activities, e.g. as a result of a neurological disease, is a serious injury to the individual. The present work aims to propose an approach to estimate the number of sessions required to learn a motor task. In the literature, there are many works on motor learning, mostly looking for ways to decrease the time of skill acquisition or motor rehabilitation. However, few works concentrate on trying to estimate the training time needed to achieve certain motor performance. In the experiment, three sessions - one session per day - were performed for each participant, whose purpose was to predict in which tracing task session the participant would reach a certain error based on their profile and initial performance. The classification models were: Algorithm K-Neighbors Nearer (KNN), Neural Network Multi Layer Perceptron (MLP), Decision Tree (AD) and Support Vector Machine (SVM). They were compared using three metrics, namely: Accuracy, F1-Score and Cohen Kappa coefficient. MLP, SVM and AD, had similar results for Accuracy and Cohen Kappa coefficient, but better than KNN, whereas for the F1-Score MPL performed better than all. This work showed the possibility of estimating the number of sessions to achieve a certain performance using prediction algorithms. This finding suggests that a similar approach may be used to estimate the amount of training a patient requires to rehabilitate.