| Summary: | This thesis investigates mechanisms of human decision making, building on the fields of psychology and computational neuroscience. I focus on human decision making measured in a psychological task with probabilistic rewards. I examine the fit of different styles of computational models to human behaviour in the task. I show that my modification to reinforcement learning, using parameters based on whether the previous trial resulted in a win or a loss, is a better fit to behaviour than my Bayesian models. Considering the task from a machine learning perspective, with the goal of gaining as many rewards as possible rather than modelling human behaviour, the performance of my modified reinforcement learning model is similar to that of my Bayesian learner and superior to that of a standard reinforcement learning model. Using population density techniques to simulate neural interactions, I confirm earlier research that demonstrates conditions which induce oscillations in a system consisting of just two nodes. I extend those findings by showing how the underlying states of the neurons contribute to complex patterns of activity. The basal ganglia form part of the brain known to be important in decision making. I create a computational model of the basal ganglia to simulate decision making. As oscillatory neural activity is known to occur in the basal ganglia, I add such activity to the model and study its impact on the decisions made. I use the time that activation first falls below a threshold as a criterion for decision making. This alternative approach allows oscillatory activity to have advantages for decision processes. Having tested my basal ganglia model on individual decisions, I extend the model to incorporate parameters related to my modified reinforcement learning model. I propose a mechanism by which the trial to trial variability observed in human responses could be implemented neurally.
|
|---|
