Multi-stage, stochastic multi-criteria decision making (MSMDM) is common in various domains of human activity, such as scientific research and combinatorial games. These decision tasks involve a series of interdependent decision-making stages where options are evaluated based on multiple criteria under uncertainty and variability. Since MSMDM is challenging for decision makers, a number of AI-powered methods have emerged to facilitate MSMDM. However, these methods have inherent limitations, e.g., dependence on training datasets and lack of transparency, especially on complex decision tasks. Therefore, we propose to explore Human-AI (HAI) collaboration approaches for MSMDM. Specifically, this thesis consists of three pieces of work, studying different HAI collaboration approaches and inv...[ Read more ]

Multi-stage, stochastic multi-criteria decision making (MSMDM) is common in various domains of human activity, such as scientific research and combinatorial games. These decision tasks involve a series of interdependent decision-making stages where options are evaluated based on multiple criteria under uncertainty and variability. Since MSMDM is challenging for decision makers, a number of AI-powered methods have emerged to facilitate MSMDM. However, these methods have inherent limitations, e.g., dependence on training datasets and lack of transparency, especially on complex decision tasks. Therefore, we propose to explore Human-AI (HAI) collaboration approaches for MSMDM. Specifically, this thesis consists of three pieces of work, studying different HAI collaboration approaches and investigating critical issues for representative MSMDM tasks. First, we take the task of deciding research directions in medicinal chemistry as our target problem and propose MedChemLens, an interactive visual system to support users to integrate the existing decision spaces and make decisions based on their various criteria. It takes an AI-assisted decision-making approach by automatically extracting and organizing molecular features from scholarly publications and visualizing the practicality of associated experiments. Second, we design RetroLens, an HAI collaborative system, which integrates two HAI collaboration methods to facilitate multi-step retrosynthetic route planning in synthetic chemistry. RetroLens adopts a joint action method to help chemists construct the decision spaces for retrosynthetic route planning together with AI and then utilizes AI-assisted decision-making to facilitate multi-criteria route revision, empowering personalized decision path exploration. Third, we focus on Go game playing and present a method, HandoverLens. This method quantifies the potential cost of assigning each decision making stage to human or AI to promote effective HAI collaboration in synchronous multi-stage decision space building and multi-criteria decision path exploration for Go playing. In all, these three pieces demonstrate the feasibility of our proposed HAI collaborative approaches to supporting MSMDM.