Examining the impact of reward landscape modulation on decision threshold selection

Abstract

The accumulation-to-threshold framework provides the best description of behaviour in perceptual decision-making tasks. Yet a long-standing question is how people select the appropriate decision thresholds. A popular hypothesis states that individuals treat sequential decisionmaking tasks as an optimisation problem, aiming to select policies which maximise reward rate. However, recent research shows that decision thresholds selected by participants are frequently
sub-optimal. Instead of abandoning the optimisation hypothesis in response to this observation, it has been suggested that one cause of sub-optimal threshold selection is the distribution of reward rate across the threshold parameter space (’the reward landscape’); that is, rather than selecting the thresholds which yield optimal reward rate, individuals mainly avoid regions of the reward landscape which yield low rewards and select their thresholds from clusters of policies that yield high, albeit sub-optimal reward rate. The current project aimed to test this hypothesis, by first identifying whether it is possible to change the reward rate distribution within the reward landscape, while keeping the optimal policy the same. A systematic manipulation of task parameters of a simulated decision task revealed that manipulating the temporal or monetary penalty parameters modulated the reward rate distribution around the optimal policy. In a subsequent experiment, the monetary penalty levels in a decision task were manipulated, to test whether changing the reward landscape would affect the policies chosen by participants.
In line with the expectations, decision-makers employed sub-optimal decision thresholds; however, there was no evidence of systematic threshold modulation as a function of reward rate distribution. Additionally, participants adopted thresholds further away from the optimal policy in a condition where this deviation caused the greatest loss of reward rate. These results suggest that decision-makers do not systematically explore the reward landscape when selecting a decision policy

Date of Award	25 Oct 2021
Original language	English
Awarding Institution	University of Bristol
Supervisor	Casimir J H Ludwig (Supervisor) & Gaurav Malhotra (Supervisor)