| dc.description.abstract | Everywhere around us machines are handed more responsibilities, because with their unique abilities, they are able to outperform humans on many tasks. However, since humans have their own unique abilities in which they still outperform machines, a logical step would be for humans and machines to collaborate in human-agent teams. For such collaboration, it is essential that they communicate smoothly. Communicating as a team member is a difficult challenge, as it requires both humans and agents to be context-sensitive and proactive.
In this thesis, an attempt was made at developing agents that learn how to communicate proactively. A combination of both data-driven learning methods as well as rule-based agent technology was used, while carefully reflecting on the influences of both methods on the learning process and the resulting behavior. Different kinds of learning agents were evaluated both in simulation as well as in an experiment where they worked together with humans. Agents learned to communicate proactively reasonably well after training in simulation, being able to use a minimal amount of communication to improve their performance as much as possible. In transferring to a context in which they played with humans, they were able to use the behaviors learned during simulation, while also learning some additional behaviors. The human team members generally trusted their agent companions, while differences in the extent to which people felt they collaborated with the agent as a team and usability were found for different kinds of communication.
This work is an attempt at bridging the different kinds of research that exist into team communication, by looking into computational and technical methods for developing proactive agent communication while constantly keeping human team members in mind. | |
