Fig. 4: Effect of learning rate, memory and sensory noise: Model captures degraded learning, non-learning, and improving with guided experience.
From: Exploration-based learning of a stabilizing controller predicts locomotor adaptation
a Stabilizing feedback controller alone only captures the fast learning transient. Addition of the reinforcement learner is needed to capture the slow transients. b Increasing the learning rate parameter speeds up learning (for a range of learning rates). c Progress toward memory makes de-adaptation faster than adaptation. d Increasing sensory noise degrades learning for fixed learning rate and fixed exploratory noise, resulting in less learning and less energy reduction. Split-belt adaptation phases are denoted by green shaded region in panels a–d. e Model captures experimental phenomena2,8 wherein a human does not adapt to an exoskeleton that provides step-frequency-dependent assistance upon the first encounter, but adapts toward the energy optimal frequency when provided with broad experience across a range of frequencies via a metronome-tracking condition. On the right two panels, blue indicates baseline condition without any assistance, red indicates exoskeleton assistance condition, and green indicates metronome-tracking condition in addition to exoskeleton assistance. In the rightmost panels, the `exo on' condition (red) shows no adaptation before the broad experience (green), but shows adaptation after the broad experience. All quantities are non-dimensional.
