Towards agent negotiators

Abstract

Negotiation is a fundamental aspect of social interaction. Due to the fundamental role it plays in our lives, there has been much work on computationally modeling negotiation. Some of the work seeks to build automated agents that can negotiate \emph{for} humans. Other work is concerned with building systems that train people to be more effective negotiators by having them negotiate \emph{against} agent negotiators that behave like humans. Whether agents are negotiating for or against humans, the central challenge is to model how humans negotiate, given the complexity of actual human behavior. This work approaches this modeling challenge by aiming to understand how people negotiate. Specifically, we ask the question of what tactics tend to occur in a single negotiation as part of a person's overall strategy and what particular combination of tactics are successful. We call these combination of tactics negotiation styles. To answer this broad question, we narrow it down to the following three sub-goals: a) identifying behavior patterns that are associated with the negotiation tactics people use (\ie tactic-related features), b) evaluating tactics effectiveness in terms of outcome and c) deriving negotiation styles (\ie co-occurrence of tactics) and evaluating the style effectiveness in terms of outcome. The negotiation studied in this work is a \textit{bilateral, multi-level, multi-issue, sequential bargaining task}, where each negotiator barters over multiple issues with varying degrees of value through a series of offers and counter-offers. Furthermore, this work takes steps towards modeling human negotiation behavior in autonomous agents using a data-driven approach by assessing data on people negotiating against a range of agents under the aforementioned bilateral negotiation task. My preliminary work demonstrated the potential of using a data-driven approach to understand how people negotiate. However, the preliminary work also identified limitations. Specifically, the dataset was collected by having people negotiate against an automated agent with a fixed response. The results were not rich in the sense that the fixed agent constrained the tactics people would use and also led to similar outcomes, making it difficult to assess how effective a tactic was. Additionally, the analysis was focused on predicting standard trait characteristics of negotiators based on tactics they used as opposed to analyzing the effectiveness of those tactics. This work goes on to address the aforementioned limitations by assessing data that is richer, specifically involving negotiations with a range of agents using different contingent response strategies. In addition, this latter work leverages different machine learning models to better characterize people's negotiation tactics and styles, as well as their effectiveness. In particular, a machine learning framework is devised coupled with an annotation schema that is informed by relevant theoretical research. This framework is applied to the aforementioned dataset to automatically produce tactic-related features that realize negotiation tactics from sub-sequences of the negotiation sequences. The produced features are further clustered into negotiation styles. The evaluation of the effectiveness of both negotiation tactics and styles are measured in terms of their capabilities to predict the negotiation outcome. By addressing the research questions, this work produces insights about people's tactics and styles in negotiation, as well as identifying the most effective tactics and styles under particular circumstances. The generated knowledge, in turn, would serve as a knowledge base for modeling human behavior in automated, autonomous agents for use in aiding or training human negotiation behavior.