Human-AI Interfaces for Operations Management
Park Sinchaisri, Professor
Business, Haas School
Applications for Spring 2024 are closed for this project.
[ To be considered, please also complete this form: https://bit.ly/45sjUZw ]
Our lab is currently working on multiple research projects of various stages on the topics of human-AI interfaces for various applications in operations management. The overarching goals are to (i) understand how humans learn and make decisions in complex environments, (ii) design interpretable algorithms to help improve human decision-making, and (iii) apply our frameworks to real-world settings. In other words, we would like to design algorithms that are useful to humans not only in theory but also in practice, taking into account interpretability/explainability, human biases, and potential aversion to comply. See my working paper as an example: https://bit.ly/tipspaper
Active projects for the 2023-2024 academic year:
1A: Human-Centered AI for Sequential Decision-Making under Uncertainty
We study how human users respond to adaptive machine-generated recommendations when they face uncertainty about the environment. We focus on an EV vehicle charging context in which drivers need to make trade-offs between charging up more frequently to reduce getting stuck in traffic and running out of charge later versus saving time and rushing through traffic. We have already designed the game and conducted a pilot experiment, so the main focus would be analyzing pilot data, improving the game design, and developing the adaptive recommendation algorithm. Current research questions: How does humans’ trust in AI evolve over time? How are uncertainty (such as traffic) and experience affect how humans interact with AI? How to optimize the delivery of machine-generated recommendations to improve compliance and performance?
1B: AI for Optimal Team Formation
AI can often suggest optimal behavior in a structured task environment, but faces the challenge of human users who may not understand how to operationalize the suggestion into their existing behavior. Collaborating with other human users could help improve their understanding and compliance with machine-generated advice. We study how individual human decision-makers exchange information/advice with one another, update their strategy after the exchange, and how to match a pair of decision-makers to form an effective team. We designed a card elimination game that tests humans’ sequential decision-making, so the main focus would be further designing the specifics of the game, modeling human behavior, and running an experiment. Current research questions: How to measure the distance between individual decision-makers’ strategies? How do humans update their strategy after exchanging information with others? How does a team of decision-makers reach a consensus or mutual strategy? What types of matches would best improve human performance? How collaboration with another human can improve the effectiveness of human-AI interfaces?
1C: AI for Improving Human-Human Collaboration
Computationally mediated work today is characterized by simple repetitive tasks carried out in isolation and mostly devoid of learning and collaboration. AI could help workers with no prior task experience to learn how to collaborate effectively and improve their skills to match their peers. We study how to model human-human collaboration as sequential decision-making (for the team) and how we can design an algorithm to help improve collaboration. We designed a two-player game in which users traverse through a grid to collect items to satisfy the mutual team goal as well as their individual goals. The main focus would be further designing the specifics of the game, modeling human-human collaboration, and running an experiment. Current research questions: How do humans collaborate (e.g., delegate tasks, share information, communicate)? What can AI learn about human collaboration and propose strategies for improvement?
1D: Human-Centered AI for Pricing
We study how to help human managers overcome biases and make optimal pricing decisions for multiple products in a retail setting (e.g., grocery store) in which the algorithm might have not observed full information that is only available to the managers. The main focus would be building up the virtual grocery store setting and running an experiment.
Current research questions: How to design an algorithm that is robust when there is uncertain information? How do humans respond to recommendations of different levels of detail?
1E: Open Topics
These are all in super early stages. Potential research questions include how to learn human advice-giving and -taking styles, how to overcome algorithm aversion, how to quantify human intuition and (counter)intuitiveness of information, how to improve human-AI collaboration, and how to best use generative AI in service settings. Open to new research questions that the students are interested in.
Role: You will gain exposure to research in human-computer interactions and behavioral operations management with real-world applications. Common stages of our projects: (i) build a strong foundation with a comprehensive literature review, (ii) design a game/setting that mimics real-world decision-making tasks, (iii) collect data via online experiments, (iv) analyze data to understand human behavior / use reinforcement learning to uncover human strategy, (v) develop interventions/algorithms to help humans improve their decision-making, and (vi) demonstrate their effectiveness via follow-up experiments or simulations. For more specific tasks, see the descriptions of projects above. Our lab meets weekly and you will get to collaborate and connect with other members in person and via our lab Discord.
Qualifications: Applicants majoring in EECS, Computer Science, Data Science, and IEOR with an intention of pursuing graduate studies (especially PhD) and/or continuing in the Spring/Summer of 2024 will be given priority.
Useful courses/skills: CS188, CS189, Javascript, Python, R, machine learning (especially reinforcement learning), front-end development.
Our most successful lab members and/or those who have had the best experience tend to be ones who are genuinely excited about the research questions, highly attentive to detail, enjoy thinking outside the box, and can be independent and self-initiative. Our lab alumni have gone to graduate programs at Harvard, MIT, Berkeley, UPenn, CMU, and UMichigan.
Hours: 9-11 hrs
Off-Campus Research Site: Open to remote as well.
Related website: https://parksinchaisri.github.io/
Related website: https://bit.ly/tipspaper
Digital Humanities and Data Science, Social Sciences, Mathematical and Physical Sciences, Engineering, Design & Technologies