Kiyofumi Miyoshi¹ (), Yosuke Sakamoto²; ¹Kyoto University, ²Osaka University

Our perceptual experience involves making decisions from a range of alternatives, and the development of comprehensive decision-making theories would greatly benefit relevant fields of vision science. Overcoming a notable hurdle in this pursuit involves addressing non-normative irrationalities that frequently emerge in decisions involving multiple options. This study specifically examines behavioral irrationalities in three-alternative dot-numerosity discrimination, focusing on the violation of the independence of irrelevant alternatives axiom (IIA violation) and the phenomenon known as the dud-alternative confidence boost. The IIA violation often manifests in the relative choice rate between the strongest target stimulus and the second strongest distractor stimulus; the relative choice rate fluctuates in response to the intensity of the least strong dud stimulus, which presents a significant challenge for broad families of decision-making theories. The dud-alternative confidence boost characterizes a puzzling scenario wherein the inclusion of a third alternative in the choice lineup leads to an irrationally heightened confidence level, surpassing that observed in a simpler two-choice problem. We propose a foundation for understanding such irrational choice behaviors through pseudo-optimal sequential evidence accumulation. Our approach considers independent evidence accumulation for three alternatives, where the difference between the first- and second-place accumulators serves as the fundamental decision variable. The dynamics of this decision variable naturally explain the IIA violation without the need for additional mechanistic assumptions. Furthermore, the early development of this decision variable immediately after stimulus onset effectively captures the trend of the dud-alternative confidence boost. To further extend the applicability of our model, we have combined the proposed decision algorithm with a deep neural network, which allows for the generation of testable behavioral predictions based on input image lineups. Leveraging this strength, we present a comparative analysis of human observers and neural networks regarding non-normative multiple-choice behaviors.