Automated vs. Human Negotiation

Automated negotiation, in which software agents assume the negotiation tasks of their human users, is argued to improve the benefits of e-business transactions. Higher benefits result on one hand from reduced transaction costs due to the avoidance of human intervention, on the other hand software agents are supposed to achieve better outcomes. While the former argument is straightforward the latter lacks empirical evidence. The few studies that compare human and software agent performance in automated negotiations only come to inconclusive results. We model and simulate automated negotiation systems and compare the output of the simulation runs to outcomes of negotiation experiments with human subjects. The automated negotiation systems consist of software agents that follow concession strategies proposed in negotiation literature and appropriate protocols that allow these agents to interrupt their strategy to avoid exploitation and unfavorable agreements. The negotiation problems used in the simulations are those derived from the experiments so that outcomes of human and automated negotiation are directly comparable. The outcome dimensions considered in our analysis are the proportion of agreements, dyadic and individual performance, and fairness. Only a set of systems managed to significantly outperform human negotiations in all outcome dimensions. These systems consist of software agents, that systematically propose offers of monotonically decreasing utility and make first concession steps if the opponent reciprocated previous concessions. The protocols of these systems enable to reject unfavorable offers to avoid exploitation or unfavorable agreements without immediately aborting negotiations.