People learn and use complex sequential actions on a daily basis, despite living in a high-dimensional environment andbody. Sequential action learning is sometimes studied in cognitive psychology using button-pressing tasks such as Nissenand Bullemers (1987) serial respone time (SRT) task. However, the SRT task only measures the speed of button presses,neglecting the richand difficult to controltrajectory of the arm, which can show predictive movements and other contextualeffects. In this study, we evolve neural networks to carry out a mouse-based SRT task under conditions of differingprediction uncertainty. We replicate behaviors found in a recent human experiment, and explore ramifications for humansequence learning.

How people learn and produce sequential actions (e.g., making coffee) has been the subject of empirical and theoreticalscrutiny, for it covers most human activities. One useful distinction is between stimulus-based control, in which actionselection is driven largely by the environment, and plan-based control, which assumes learning of structured sequences ofactions and effects. Task demands, instructions, and participants’ individual abilities and inclinations can all modulate thecontrol mode used. We investigate two sequence learning tasks, with one key difference: in the cued task either controlmode is possible, while learning in the reinforcement task requires plan-based control. Using measures of visuospatialworking memory (VWM) capacity, locus of control, need for structure, and IQ, we seek to explain individual differencesin choice of control mode and task performance, establishing a link between VWM capacity and performance, as well asexplicit knowledge evidencing plan-based control.

We humans spend almost a third of our lives asleep, andthere is mounting evidence that sleep not only maintains, butactually improves many of our cognitive functions. Mem-ory consolidation–the process of crystallizing and integratingmemories into knowledge and skills–is particularly benefittedby sleep. We survey the evidence that sleep aids memory con-solidation in various declarative and implicit tasks and reviewthe basic neurophysiological structure of sleep with a focus onunderstanding what neural systems are involved. Drawing onmachine learning research, we discuss why it might be usefulfor humans–and robots, perhaps–to have such an offline pe-riod for processing, even though humans are clearly capable oflearning incrementally, online. Finally, we propose and simu-late two mechanisms for use in computational memory modelsto accomplish sleep-based consolidation via either or both 1)re-encoding knowledge representations and 2) reactivating andstrengthening recent memories.

Learning sequential actions is an essential human ability, formost daily activities are sequential. We modify the serial reac-tion time (SRT) task, originally used to teach people a con-sistent sequence of button presses by cueing them with thenext target response, to record mouse movements, collectingcontinuous response trajectories. Further, we introduce a rein-forcement learning version of the paradigm in which the nexttarget is not cued. Instead, learners must explore response al-ternatives, and receive a penalty for each incorrect response,as well as a reward for a correct response. Participants arenot told that they are to learn a single deterministic sequenceof responses, nor that it will repeat (nor how often), nor howlong it is. Given the difficulty of the task, it is unsurprisingthat some learners performed poorly. However, many learn-ers performed remarkably well, and some acquired the full 10-item sequence within 10 repetitions. We compare the high- andlow-performers’ detailed results in this reinforcement learning(RL) task with a cued trajectory SRT task, finding both simi-larities and discrepancies. Finally, we note that humans in thistask outperform three standard RL models and have differentpatterns of errors that suggest future modeling directions.