Introduction
Automated negotiation, where software agents bargain on behalf of human interests, is a cornerstone of multi-agent systems research. However, researchers often face a trade-off between legacy Java-based frameworks like GENIUS and modern Python tools like NegMAS, which may lack specialized analytics for opponent modeling and "negotiation dance" efficacy. NegoLog addresses this gap by providing a Python-based framework that streamlines agent development while offering granular insights into negotiation dynamics.
The NegoLog Architecture
NegoLog is built upon the modular BOA (Bidding strategy, Opponent model, Acceptance strategy) architecture, primarily utilizing the Stacked Alternating Offers Protocol (SAOP). A defining feature of the framework is its ability to evaluate opponent preference models independently of bidding strategies.
As shown in the architecture above, NegoLog allows for "listener" models that estimate preferences in real-time without influencing the actual bids. This decoupling enables researchers to determine whether an agent's failure stems from its bidding strategy or an inaccurate model of the opponent, facilitating more effective debugging of agent "intelligence."
Advanced Assessment and Metrics
The framework automates the calculation of complex metrics that are often manually implemented by researchers. This includes statistical correlations (Spearman and Kendall-Tau) and distance metrics (Nash and Kalai-Smorodinsky) to evaluate deal quality relative to theoretical optima.
NegoLog categorizes its analytics into three primary domains: Session Process, Negotiation Outcome, and Opponent Model Estimation. By providing built-in loggers for these categories, the framework facilitates a deep-dive analysis of the "negotiation dance"—the trajectory of offers in the utility space—to identify if agents are successfully finding Pareto-optimal solutions or being unnecessarily selfish.
Domain Generation and Accessibility
To support rigorous benchmarking, NegoLog includes a Domain Generator that creates synthetic scenarios based on "balance scores" (fairness) and "opposition levels" (conflict). It remains compatible with the Automated Negotiating Agents Competition (ANAC) repositories, allowing for head-to-head comparisons against established agents. Furthermore, a web-based user interface lowers the barrier to entry, replacing complex configuration files with an intuitive tournament management system.
Conclusion and Future Directions
While NegoLog currently focuses on bilateral negotiations, it provides a robust foundation for evaluating automated negotiators through its unique assessment components. Future development aims to extend support to multilateral bargaining and integrate reinforcement learning hooks for self-play training environments.
Summary updated