Working memory in reinforcement learning

Abstract

The ability of an individual to learn from delayed action outcomes helps them to better adapt to their environment. In order for learning to take place, they might utilize working memory to hold in mind a response they have just made and compare it to feedback received. This dissertation encompasses three experiments that test several putative neural correlates of working memory during the prefeedback delay. In the first experiment, I show that two EEG measures, the stimulus-preceding negativity (SPN) and lateralized theta, decline in amplitude over the course of learning a new motor skill. In the second experiment, I show that SPN and frontocentral beta are of greater amplitude in a condition that imposes a high load on working memory for actions. In the third experiment, I show that SPN and frontocentral beta increased progressively across conditions with one, two, or three cues in a probabilistic categorization task. This effect was greatly reduced in participants whose ability to develop complex hypotheses regarding the relation of stimuli, responses, and outcomes was impaired by Parkinson's Disease. By adopting three different approaches to testing the relationship of EEG measures to the brain's ability to bridge the gap between responses and outcomes, this dissertation provides converging evidence regarding the role of working memory.