Prediction Mode

In prediction mode, the system has already learned with CEM and can therefore use the incoming data to forecast a small step into the future. This situation is depicted in Figure 8.2. Here, the output motion of user B is actually being generated by the learning system instead of directly piping out of his vision system. It becomes quickly apparent that nothing too interesting can happen as a result of this mode of operation. The ARL system is simply operating as a filter (i.e. a Kalman filter) since it is only generating some slightly modified version, ${\hat {\bf y}}_B(t)$ of the original signal ${\bf y}_B(t)$. When user B exits the scene, the system merely locks up since the ${\bf x}(t)$ is only being fed half of the signal (from user A). The ${\bf x}$ vector only contains memory about user A and the system has no memory of its own actions. Thus, only a portion of ${\bf x}$ is being updated and poor estimates using the pdf $p({\bf y}\vert{\bf x})$ will result. Thus, this mode of interaction is merely a filtering where the output gestures of user B just seem smoothed.

  

Figure 8.2: Prediction Mode