影响对人工智能决策支持系统适当依赖的因素

iiiACKNOWLEDGEMENTSI want to acknowledge the incredible mentorship, coaching, and support of myadvisorDr. Baruch Fischhoff, and the other members of my committee, Dr. Peter Adams, Dr. CaseyCanfield, and Dr. Stephen Broomell. I also appreciate collaborating over several years with Dr.Alex Davis in producing this research. Dr. Davis’s love for research, dedication to details, andcommitment to student development set the foundation for this work. I am especially grateful toMartin Liu who diligently worked on the development of the experimental platform while as agraduate student in Heinz, during the COVID-19 pandemic, and after graduation. I want to thankthe Swedish Foundation for Humanities and Social Science for partial funding support. Thiseducation would not have been possible had it not been for the incredible support,professionalism, and funding through the Army’s Strategic Policy and Planning Program(ASP3). I am grateful to Dr. Robert Davis, COL(R) Chris Prigge, and the ASP3 support staff,without whom I would not have been able to tackle this endeavor. Finally, and most importantly,I owe an immense amount of gratitude to my wife, Marina Shin, for her love, patience, support,and dedication as well as to our children, Daniel and Emily Shin, for their constant encouragement. ivABSTRACTMany applications of AI require humans and AI advisors to make decisionscollaboratively; however, success depends on how appropriately humans rely on the AI agent.We demonstrated an evaluation method for a platform that used neural network agents of varyingskill levels for the simple strategic game of Connect Four. We manipulated the presence,sequence, skill, and information display of Artificial Intelligence (AI) advice in a strategy gameagainst another AI opponent that sometimes varied its skill to measure their effect on users’Human agent teams outperformed unaided subjects with those receiving the AIrecommendations simultaneously achieving the best results. Although team performance washigher and subjects improved during game play, there was little evidence of learning from theirAI advisors. AI reliability proved to be the greatest determiner of team performance withsubjects retaining trust in higher skilled advisors even in varied environments. Those with highernumeracy demonstrated the highest ability to make use of AI advice including more detailedoutput formats including ranking of choices and probabilities.More reliable AI agents correlatedto higher AItrust while higherself-confidencecorrelated togreaterrejection of AI advice,greater confidence in success, but slightly lower performance.The value of these human agent teams depended on AI reliability, users’ ability to extractlessons from their advice, and users’ trust in that advice. Organizations implementing humanagent teams should conduct testing to know how well users appropriatelyrely on AI performance.recommendations. vTABLE OF CONTENTSABSTRACT……………………………………………………………..……………….ivLIST OF TABLES………………………………………………………………………..viLIST OF FIGURES……………………………………………………...……………….xx1.Introduction…………………………………………………………..……………12.When Do Humans Heed AI Agents’ Advice? When Should They?…………….…63.How Well Do Human Decision Makers Use of AI Agents' Advice? Does DisclosingAdvisor Uncertainty Help?…………………………………………….…………354.Losing and Gaining Faith Using An Artificial Intelligence Decision SupportSystem When The Situation Changes …………………………….……………...645.Conclusion……………………………………………………………..……….1076.Supplemental Material………………………………………..………………...1187.Appendix A, Experiment 1………………………………….……………….….1808.Appendix B, Experiment 2…………………………..………………………….2019.Appendix C, Experiment3……………………………………………………...224REFERENCES…………………………………………………………………………271 Chapter List of Tables1.Table 1. Artificial Intelligence (AI) agent characteristics…………………………………….102.Table 2. Win, loss, and draw rates for all 7 experimentalgroups………………………….….303.Table 3. Sequence of play and type of display descriptive statistics…………………….……474.Table 4. Win, loss, and draw rates for all 7 experimental groups……………………………..595.Table 5. Artificial Intelligence (AI) agent characteristics…………….……..………………..676.Table 6. Win, loss, and draw rates for all 4 experimental groups……………………………..787.Table 7.Conditions, results, and meaning of human decisions incorporating AI advisorrecommendationsbased on skill score, choice selections, provisional choices, and AIadvice ……………………………………………………………………………………..…818.Table 8. Win, loss, and draw rates (and standard deviation) for all 4 experimental groups duringfree play without an AI advisor and the control group from experiment #2………………….939.Table SM 1.Artificial Intelligence (AI) agent characteristics recorded during development andthe experiment…………………………………………………………………….………...12110.Table SM 2.Table SM2. Win, loss, and draw rates for all 7 experimental groups……........13111.Table SM3. Artificial Intelligence (AI) agent characteristics recorded during development andthe experiment……………………………………………