As emerging human-machine interfaces enable new opportunities for enhanced performance, users must acquire appropriate motor skills to operate the technology through lengthy, intensive, and elaborate training. The interfacing system capable of stimulating the processes of learning is key to minimizing the training-related costs and improving the effectiveness of human performance. Model-based frameworks can enable optimal training through computational modeling to better understand the process of motor learning and simulations to explore the training conditions that facilitate skill acquisitions. In this dissertation, we conducted reaching experiments on 15 healthy individuals and explored the role of error-based feedback and its augmentations (error-augmentation or EA) to shape the time course of motor learning. Computational models tested several learning mechanisms including 1) constant versus error-dependent learning rates, 2) first-order versus higher-order processes, and 3) generalization across movement directions.

We first determined the sample size of trial performances needed to best assess the time course of motor learning using Monte-Carlo simulations, and we found that the number of samples required increased linearly with the time constant of transient (learning) signals and decreased exponentially with the signal-to-noise ratio. Next, we found that the second-order proxy process model with the trial-to-trial update method best predicted the influence of EA (gain and offset) on changes in movement errors from sparse, intermittent no-vision (catch) trials, which are generally accepted as truer indicators of the formation of the internal models. Next, our simulations using this cross-validated model revealed that the optimal EA (gain and offset) schedules should be held constant near 2 to minimize the final error and maximize the rate of learning, and the training duration of at least 16 trials was necessary. We then identified generative model structures to predict motor learning through randomized reaching practice across movement directions, and we found that iterative update of the initial ballistic launch to reaching targets was predictable using the direction-specific, locally generalizing first-order model with constant learning rate. Finally, we investigated generative model structures that can predict motor learning through corrective submovements, and we found that iterative update of the submovements during rapid reaching was predictable using the first-order affine model with Gaussian-weighted learning rate.

We determined sample sizes needed to best assess learning curves, identified simpler phenomenological models to predict the influence of EA on learning, and identified intricate generative models that can predict updates to the reaching movements during learning. In the field of neurorehabilitation, robotics and virtual reality-based systems can deliver therapy at a much higher intensity and dosage through novel interactivity that stimulate neuroplasticity. However, such therapy approaches have achieved comparable if not only modest gains over conventional therapy. This is where model-based frameworks such as the ones presented in this dissertation have the potential to provide diagnostic insights into impairment characteristics of patients and enable therapists to explore treatments that generate desired clinical outcomes. This can also dramatically impact sports, piloting, and other forms of performance training.