A comparison of two agent interaction design approaches

The approach to designing agent interactions that is used by mainstream agent-oriented software engineering methodologies focuses on identifying the allowable sequences of messages, and capturing this as an interaction protocol. It has been argued that this "message-centric" approach is not congruent with the ability of individual agents to persistently achieve goals in a flexible and robust manner. In this paper we report on an empirical comparison of a message-centric approach to designing interactions exemplified by Prometheus and a previously proposed alternative approach called "Hermes" that uses interaction goals. The empirical comparison had 13~participants, each of whom created a design for the agent interactions in a meeting manager system. Six of the participants used Hermes, the remaining seven used Prometheus. The designs produced were analysed to assess their performance against a range of criteria including flexibility number of pathways, the degree to which they covered the provided scenario, and robustness ability to deal with a range of pre-defined exceptional behaviours. We also measured the time taken to develop the design, and surveyed participants to assess their opinions on their designs. The comparison showed that Prometheus did indeed lead participants to develop designs that had significantly less flexibility and robustness than Hermes, and that the designs of the Hermes group did significantly better at covering the scenario. On the other hand, Prometheus was significantly faster to use. The survey responses did not display a statistically significant difference, with the exception that Prometheus users felt significantly more strongly than Hermes users that their design was easy to follow.