Handing over control: When it is better for an arti cial agent to let the human decide?
This thesis considers human-robot collaborative scenarios, where the agent can be in need of assistance from a human collaborator. As for all autonomous systems, performance is very important. In addition to performance, it is also desirable that a human-agent collaborative system does not interrupt the human continuously: It needs to be able to pick good moments to do so, in order to not interrupt the human collaborator, while maintaining a good level of performance. A collaborative chess game was implemented and used as an instrument to conduct experiments. In this game, two agents play against each other. One of these can decide to ask for help from an expert agent, based the distribution of evaluation values given a state. Using this framework, the research questions can be researched, pertaining to the level of per- formance of asking for help versus not doing so and whether the timing matters: We test this by comparing the performance of the former to an agent that hands over control at random intervals, but at the same fre- quency. Experimental results showed that help-asking frequency in uenced per- formance and that the timing at which these help-calls were uttered were important to the performance as well.
Faculteit der Sociale Wetenschappen