MoCaNA, un agent de n{e}gociation automatique utilisant la recherche arborescente de Monte-Carlo

Automated negotiation is a rising topic in Artificial Intelligence research. Monte Carlo methods have got increasing interest, in particular since they have been used with success on games with high branching factor such as go.In this paper, we describe an Monte Carlo Negotiating Agent (MoCaNA) whose bidding strategy relies on Monte Carlo Tree Search. We provide our agent with opponent modeling tehcniques for bidding strtaegy and utility. MoCaNA can negotiate on continuous negotiating domains and in a context where no bound has been specified. We confront MoCaNA and the finalists of ANAC 2014 and a RandomWalker on different negotiation domains. Our agent ouperforms the RandomWalker in a domain without bound and the majority of the ANAC finalists in a domain with a bound.

💡 Research Summary

The paper introduces MoCaNA, a novel automated negotiation agent that leverages Monte‑Carlo Tree Search (MCTS) for its bidding strategy and incorporates sophisticated opponent modeling techniques. The authors first formalize bilateral bargaining as an extensive‑form game in which two agents alternately exchange offers until an agreement is reached, a rejection occurs, or a predefined termination condition is met. Each offer consists of a set of continuous attributes, allowing the agent to operate in domains without discretization or explicit time/round bounds.

MoCaNA’s core consists of four MCTS phases. In the selection phase, a UCT‑based policy balances exploitation of high‑value nodes with exploration of under‑sampled branches, ensuring that novel offers are examined early. During expansion, the agent samples new offers from the continuous space using Gaussian Process Regression (GPR), which predicts promising attribute values based on the opponent’s historical proposals. The simulation phase runs a “play‑out” of the negotiation using a learned opponent model: GPR predicts the opponent’s next offer, while a Bayesian utility estimator infers the opponent’s preference weights over attributes (linear or piece‑wise linear). This dual model enables realistic prediction of acceptance or rejection. In the back‑propagation step, the simulated outcome’s reward—computed as a weighted combination of the agent’s own utility and the estimated opponent utility—is propagated up the tree, guiding future search toward mutually beneficial agreements.

The authors evaluate MoCaNA against the six finalists of the ANAC 2014 competition and a baseline RandomWalker in two settings: (1) an unbounded domain with no round limit, and (2) a bounded domain with a fixed number of rounds. Results show that MoCaNA consistently outperforms the RandomWalker in the unbounded scenario and surpasses the majority of ANAC finalists in the bounded scenario, demonstrating the effectiveness of MCTS combined with opponent modeling even when the negotiation space is large and continuous.

Key contributions include: (i) a formal mapping of bilateral bargaining to an extensive‑form game suitable for MCTS, (ii) the integration of GPR for continuous opponent offer prediction, (iii) a Bayesian approach to infer opponent utility without prior type distributions, and (iv) empirical evidence that MCTS can handle negotiation problems traditionally considered unsuitable for tree‑search due to high branching factors and incomplete information.

Limitations are acknowledged: the computational cost grows with the number of simulations, early‑stage inaccuracies in opponent models can misguide the search, and the current design is limited to two‑party negotiations. Future work is suggested on scaling to multi‑party settings, optimizing simulation budgets, and exploring alternative tree‑expansion heuristics.

Overall, MoCaNA represents a significant step forward in automated negotiation research, showing that techniques pioneered in game AI—particularly Monte‑Carlo tree search—can be successfully adapted to complex, continuous, and potentially unbounded negotiation environments.