An opponent model for agent-based shared decision-making via a genetic algorithm

Kai-Biao Lin, Weiqin Ying, Yong Liu, Fei-Ping Hong, Yimin Yang, Ping LüFrontiers in Psychology 2023 · DOI

Abstract

The agent negotiation framework supports Shared Decision-Making (SDM) participants in accessing treatment plans that fit individual preferences, thereby increasing treatment satisfaction. Adding Genetic Algorithm (GA) opponent preference prediction to the SDM negotiation framework can effectively improve negotiation performance in incomplete information environments.

Introduction

Shared Decision-Making (SDM) is considered the "gold standard" of modern healthcare, where doctors and patients collaborate to choose a treatment plan. However, the reality is often complex. Doctors are under extreme time pressure, and patients may lack the medical literacy to clearly articulate their priorities, such as valuing side-effect avoidance over total cost. This creates an incomplete information environment where negotiations can stall or lead to suboptimal decisions.

To address this, researchers have developed the Agent-based Negotiation framework using Fuzzy constraints and a Genetic Algorithm (ANFGA). This system automates the negotiation process using intelligent agents that can predict the other party's hidden preferences to reach a win-win agreement efficiently.

The ANFGA Framework

The ANFGA system consists of two primary agents: a Doctor Agent (DA) and a Patient Agent (PA). The framework utilizes several key components to manage the negotiation:

  1. Fuzzy Constraint Satisfaction: Instead of rigid preferences, the model uses trapezoidal fuzzy membership functions. This allows agents to represent varying satisfaction levels across a range of values (e.g., varying degrees of "happiness" for a treatment cost between $2,000 and $6,000).
  2. Stacked Alternating Offers Protocol: The agents take turns making offers and counter-offers in a structured cycle.
  3. Concession Strategies: Agents can be tuned to different behaviors—Competitive, Collaborative, or Win-Win—adjusting how quickly they lower expectations to reach a deal.
The process of ANFGA with a PA first offer.

Genetic Algorithm for Opponent Modeling

The core innovation of ANFGA is the use of a Genetic Algorithm (GA) to model the opponent's preferences. In a typical clinical negotiation, one agent doesn't know the other's internal "weight" for different issues (like cost vs. efficacy).

The GA maintains a population of potential preference profiles. As the negotiation progresses, the GA "evolves" these profiles, using the history of the opponent's offers to calculate "fitness." Over time, the GA converges on a model that accurately predicts what the opponent actually values. This allows the agent to offer counter-proposals that are more likely to be accepted without sacrificing its own core requirements.

Opponent preference prediction process using GA.

Key Results and Insights

The researchers compared ANFGA against several benchmarks, including models without opponent prediction (FCAN) and purely time-driven models (ANF-TIME). The study utilized real-world preference data from the Department of Pediatrics at Xiamen Hospital of Traditional Chinese Medicine alongside expanded simulations.

  • Success Rate: ANFGA-Collaboration and ANF-Win-Win maintained a 100% Negotiation Success Rate (NSR) across almost all deadlines and issue complexities.
  • Satisfaction: ANFGA reached higher Average Joint Satisfaction (AJS) scores than non-predictive models. In a 5-issue scenario, ANFGA achieved satisfaction levels nearly identical to a "Complete Information" environment where all preferences are known upfront.
  • Efficiency: By learning that a patient might care deeply about "Side Effects" but be flexible on "Cost," the Doctor Agent can reach an agreement much faster, effectively mitigating the time pressure found in real-world clinical settings.

Conclusion

ANFGA demonstrates that predictive modeling is the "secret sauce" for successful automated negotiation in healthcare. By using Genetic Algorithms to bridge the information gap between doctors and patients, the system ensures more robust and satisfying treatment plans. Future work will focus on testing the model in real clinical environments to see how it handles the inherent unpredictability of human behavior.

Summary updated

Abstract

The agent negotiation framework supports SDM participants in accessing treatment plans that fit individual preferences, thereby increasing treatment satisfaction. Adding GA opponent preference prediction to the SDM negotiation framework can effectively improve negotiation performance in incomplete information environments.
BibTeX
@inproceedings{lin2023opponent,
  author = {Kai-Biao Lin and Weiqin Ying and Yong Liu and Fei-Ping Hong and Yimin Yang and Ping Lü},
  title = {{An opponent model for agent-based shared decision-making via a genetic algorithm}},
  booktitle = {Frontiers in Psychology},
  year = {2023},
  doi = {10.3389/fpsyg.2023.1124734},
}
Figure