The Effectiveness of Competitive Agent Strategy in Human-Agent Negotiation

Human-agent negotiation is a social task that provides a multifaceted proving ground for artificial intelligence systems that aim to interact with humans in a social context. Designing agents that are capable of negotiating with humans provides threefold benefit. First, it allows information regarding human behavior to be gleaned in an efficient and repeatable context through the use of programmable agents, which can serve as perfectly consistent and customizable confederates in empirical studies. Second, these agents are allowed to be tested in a real-world context, and theoretical strategies and behaviors that make the agents more effective are able to be refined directly. Finally, the agents are able to provide feedback for their human partners, directly improving their negotiation abilities and providing personal benefit to the study participants. This work demonstrates the results of a study conducted on the Interactive Arbitration Guide Online (IAGO) Negotiation platform. The study compares the effectiveness of four different types of automated agents as they negotiate with humans over the course of a 10-minute interaction. The agents differ in a 2x2 design according to agent competitiveness (competitive vs. consensus-building) and agent attitude (nice vs. nasty attitude). These results show that in this multi-issue bargaining task, competitive agents performed far better than consensus-building agents against their human opponents, scoring far more points than the humans did. In contrast to some previous work, there was not a significant effect of agent attitude. These results have impact on agent design for single, one-shot interactions resembling real-world negotiation, although they may not extend to repeated interactions. Author