Understanding Human Choices as Computationally Rational Processes

Chen, Haiyang (2022). Understanding Human Choices as Computationally Rational Processes. University of Birmingham. Ph.D.

Preview

Chen2022PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (3MB) | Preview

Abstract

Risky choice involves deciding between gambles that can differ in the probability and value of outcomes. This thesis exposes the cognitive processes that underpin risky choice in humans. The approach taken involves the use of deep neural networks and reinforcement learning to discover policies that are adaptive to distributions of risky choice problems. Risky choice has been extensively studied for hundreds of years and in the modern era many phenomena have been reported. Sometimes these phenomena are explained away as ``irrationalities'' or ``biases''. This thesis uses computational methods to demonstrate that apparently irrational risky choice can be ecological rational and sometimes rational given cognitive bounds. Moreover it does so for a broader range of risky choice problems than has so far been investigated. These include both contextual choice problems and the fourfold pattern of risky choice. The implications for future work are discussed.

The results show that (1) context effects (including attraction, compromise and similarity) can emerge from an optimal (rational) ``classifier'' that chooses the option with the highest expected value; (2) the new model could predict context effects, as for people, when the representation format encourages attribute comparisons; (3) the new model approximates a bounded optimal cognitive policy and makes quantitative predictions that correspond well to evidence about human contextual choice; (4) an alternative explanation that a wide range of risky choice phenomena emerge from boundedly optimal adaptation of a decision making agent to processing constraints. In each study, the model is not pre-programmed to process all information but learns to process only that information that helps it maximize utility. We argue that the models provide evidence that apparently irrational risky choices are emergent consequences of processes that prefer higher value (rational) policies or classifiers.

My thesis is that a number of models offer novel and rational explanations for a broad range of phenomena exhibited by people making choice under risk. I demonstrate that apparent cognitive biases can emerge from computational rational processing. Furthermore, I propose a unifying framework for modelling risky choice phenomena. Deep reinforcement learning has the potential to help discriminate between various explanations because it provides a means of computing computationally rational policies given both ecological and cognitive bounds.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):