CaltechTHESIS
  A Caltech Library Service

Attention, Strategy, and the Human Mind

Citation

Li, Xiaomin (2021) Attention, Strategy, and the Human Mind. Dissertation (Ph.D.), Devision of Humanity and Social Science. doi:10.7907/6zqy-pt73. https://resolver.caltech.edu/CaltechTHESIS:11112020-185040796

Abstract

The current dissertation chapters try to discover the role of visual attention in decision making from three different perspectives: 1) how attention bias affects strategic decision makings, 2) how to model eye movement data to better understand strategic decisions, 3) how to manipulate simple choices through visual saliency.

The second chapter introduces a series of novel image games, where players need to match, hide, or seek against other players. We apply a pure computational way: the state-of-art visual saliency algorithm, Saliency Attentive Map (SAM) to measure visual saliency. We find that visual saliency can predict strategic behaviors well. The concentration of salience is correlated with the rate of matching when players are both trying to match location choices (r=.64). In hider-seeker games, all players choose salient locations more often than predicted in equilibrium, creating a ``seeker’s advantage'' (seekers win 9\% of games rather than the 7\% predicted in equilibrium). The 9\% win rate is robust for paying higher stakes and using a between-subjects design. Salience-choice relations are consistent with cognitive hierarchy and level-k models in which strategically naive level 0's are biased toward salience, and higher-level types are not directly biased toward salience, but choose salient locations because they believe lower-level types do. Other links between salience as understood in psychology and hypothesized in economics are suggested.

The third chapter is a continuation of the second chapter, but with a different emphasis. The third chapter proposes a way to dynamically model gaze transitional data in games utilizing a class of machine learning model: hidden markov models(HMM). The HMM model reveals how the attentional bias affects strategies on different time point. Besides, this model well connects to the k level behavioral method and can make novel predictions on strategic levels. With further containing the fixation duration data, we developed a continuous-time hidden Markov model (cgtHMM), which can be used to predict how exactly time pressure changes choices and the seeker’s advantage.

Distinct from the other two, chapter four aims at manipulating binary choice outcomes through the change of visual saliency distribution under SAM. We design a value-based choice paradigm where both the reward property and the attention property are well separated and controlled. The experimental results indicate that visual saliency can enhance the choice correction rates when the more rewarding outcome is also labeled salient. It can also shorten the decision time needed. Such a result can be explained by a saliency-enhanced rational inattention model by incorporating attention factors in the traditional RI model.

Item Type:Thesis (Dissertation (Ph.D.))
Subject Keywords:Visual Attention, Game Theory
Degree Grantor:California Institute of Technology
Division:Humanities and Social Sciences
Major Option:Behavioral and Social Neuroscience
Thesis Availability:Public (worldwide access)
Research Advisor(s):
  • Camerer, Colin F.
Thesis Committee:
  • Tamuz, Omer (chair)
  • Adolphs, Ralph
  • Jin, Lawrence J.
  • Camerer, Colin F.
Defense Date:30 September 2020
Non-Caltech Author Email:lixiaominxiaomin (AT) gmail.com
Record Number:CaltechTHESIS:11112020-185040796
Persistent URL:https://resolver.caltech.edu/CaltechTHESIS:11112020-185040796
DOI:10.7907/6zqy-pt73
ORCID:
AuthorORCID
Li, Xiaomin0000-0002-1286-4012
Default Usage Policy:No commercial reproduction, distribution, display or performance rights in this work are provided.
ID Code:13997
Collection:CaltechTHESIS
Deposited By: Xiaomin Li
Deposited On:12 Nov 2020 18:34
Last Modified:18 Dec 2020 00:26

Thesis Files

[img] PDF - Final Version
See Usage Policy.

21MB

Repository Staff Only: item control page