USER EVALUATION
In the previous section, we described a framework for training and evaluating strategies
for social multiuser interaction in simulation. In this version, the communication
of social state updates from the Social State Recognizer (SSR) to the Social Skills Executor
(SSE) happened on a frame-by-frame basis. For the full robot bartender system,
however, the event-based SSR as described in Section 2 was used, which only publishes
new social states if a relevant change occurred. The learning framework has therefore
been adapted to incorporate this new SSR, and the SSE is hence triggered by state
update events, and—as described in more detail in Keizer et al. [2013a]—new policies
for this setup were trained using MUSE.