Electrophysiological evidence for a universal reward prediction error encoder in reinforcement learning

Stewardson, Harry (2022) Electrophysiological evidence for a universal reward prediction error encoder in reinforcement learning. Doctoral thesis, University of East Anglia.

Preview

PDF Download (4MB) | Preview

Abstract

In human decision-making, choices are made with the ultimate aim of maximising favourable outcomes while minimising unfavourable ones. This is not determined merely by an assessment of outcomes being good or bad, but by them being better or worse than anticipated before their occurrence. The difference between expected and attained outcomes can be quantified in the form of reward prediction errors, whereby positive values represent rewards and negative values instead represent punishments. Research in Neuroeconomics demonstrates a convincing argument that the human brain generates signals representative of such reward prediction errors for the purposes of reinforcement learning. Isolating such signals would stand to tell a great deal about not only typical human decision-making, but also prove beneficial for furthering the understanding of non-normative behaviour such as psychopathological gambling and addiction. The present thesis develops this research by focussing on the feedback-related negativity, an electrophysiological component largely attributed to the reflection of a neural encoding of reward prediction errors in reinforcement learning tasks. While the feedback-related negativity is the most widely researched component regarding this topic, methodological differences in its study and the intricacies of component overlap has led to substantial debate regarding the precise type of reward prediction error encoding it is said to represent. In this thesis I have broken down these types of encoding and accurately identified that most represented in the time interval of the feedback-related negativity; counter to many claims, I have demonstrated this encoding to not only be present using experimental designs that have previously led to its misidentification, but also present within the pre-existing literature following thorough meta-analytical approaches utilizing Bayesian methods. In line with these findings, I propose the feedback-related negativity as a signal representative of universal, or general-purpose, reward prediction error encoding in all manner of human reinforcement learning scenarios.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Social Sciences > School of Psychology
Depositing User:	Chris White
Date Deposited:	28 Nov 2022 08:45
Last Modified:	28 Nov 2022 08:45
URI:	https://ueaeprints.uea.ac.uk/id/eprint/89965
DOI:

Downloads

Actions (login required)

View Item