Abstract
Human decision-making often suffers from "narrow bracketing," where multiple choices are evaluated in isolation rather than as a collective portfolio. This paper presents SAMP (Selection problem Advice provider for Multiple Problems), an AI agent designed to mitigate this bias by providing advice that maximizes human satisfaction. By integrating Prospect Theory with machine learning, SAMP models subjective human preferences regarding risk and probability. Experimental results demonstrate that SAMP significantly outperforms traditional rational agents—which focus solely on expected monetary value—in terms of user satisfaction and perceived advice quality.
Introduction
When faced with multiple concurrent decisions, such as managing various financial investments or life choices, humans typically treat each task as a separate entity. This cognitive shortcut, known as "narrow bracketing," often leads to suboptimal global outcomes. While "broad bracketing" (considering all choices simultaneously) is mathematically superior, it is cognitively demanding for most individuals.
The challenge lies in developing automated advisors that can help humans bracket their choices broadly. However, a "perfectly rational" advisor that maximizes expected monetary value (EV) often fails to satisfy human users, who possess inherent subjective biases toward risk and probability. This research introduces SAMP, an agent that bridges the gap between mathematical optimization and human psychology.
Methods
The SAMP agent employs a multi-stage workflow to provide personalized advice:
- Preference Elicitation: A learning group of 52 participants provided data on simple binary choices (e.g., a guaranteed sum vs. a probabilistic win).
- Modeling via Prospect Theory: Using this data, a logistic regression classifier was trained to elicit "decision weights." These weights reflect how humans subjectively perceive probabilities, adhering to the principles of Tversky and Kahneman’s Prospect Theory.
- Optimization: For a given set of multiple problems, SAMP searches the combinatorial space of all possible choices to identify the set that yields the highest human-centric utility based on the learned model.
- Visualization: Advice is presented to the user via a pie chart, illustrating the probability distribution of potential final outcomes if the advice is followed.
Results
The performance of SAMP was evaluated against a "Rational Agent" (which maximizes Expected Value) in a study involving 202 participants. Satisfaction was measured across three dimensions on a 1–5 scale: happiness with the final result, happiness with the decisions made, and the perceived quality of the advice.
Statistical analysis (ANOVA) revealed that SAMP users reported significantly higher satisfaction ($p < 0.05$) in all categories. This suggests that an agent modeled on human psychological principles is a more effective advisor than one based on pure economic rationality.
Discussion and Conclusion
SAMP demonstrates the power of offline learning; by modeling general human preferences from a separate group, the agent can provide immediate, high-quality advice to new users without requiring individual training. This approach simplifies complex multi-variable problems into intuitive visualizations.
However, limitations remain. The current model assumes identical utility between the advisor and the user, which may not hold in commercial settings like brokerage. Additionally, the search space for optimal advice grows exponentially with the number of problems, necessitating more advanced heuristic searches for larger scales. Future research will explore real-time fine-tuning of models to individual risk tolerances.
Summary updated