Comparing The Accuracy of Frequentist and Bayesian Models in Human-Agent Negotiation

Understanding an opponent's wants is crucial for maximizing the outcomes of a multi-issue negotiation. To do this, automated systems must build an "opponent model" from information conveyed during a negotiation. Bayesian and frequentist models are the most commonly used. Bayesian models have a principled way to incorporate prior knowledge about an opponent's preferences. However, frequentist models have outperformed Bayesian approaches in practice, dominating the yearly agent-verses-agent negotiation competitions. With growing interest in agents that negotiate with people, this presumed dominance needs to be revisited. Human opponents convey far less information than automated agents, and people often share similar preferences (e.g., in a salary negotiation, most people care the most about salary). Thus, the theoretical advantage of Bayesian approaches may translate into practice for agent-versus-human negotiation. In this work, we compare the performance of Bayesian models against a leading frequentist approach in an agent-versus-human multi-issue salary negotiation. Although we show that frequentist opponent models outperform Bayesian models when using a uniform prior, Bayesian approaches excel when using two common priors. The best performance is achieved with an empirically-derived prior (i.e., biasing the model space using the distribution of preferences found in past human negotiators). Yet, strong performance is also observed when using a "fixed-pie bias", the prior used by most human negotiators. We discuss the implication of these findings for research on human-agent negotiation.