Simulation Mode

  

Figure 8.5: Simulation Mode

As shown in Figure 8.5, the predicted user's action ${\hat {\bf y}_{A}}$ could be fed back into the time series doing away with any human input altogether. Note the perfectly symmetric paths of the data flow. There are no vision systems in operation. Instead, we only see two graphics systems which depict the ARL system's actions. The actions are fed back into its short-term memory which represents the past virtual actions of both user A and user B.

Unlike the previous operation mode which had at least one human generating a component of the signal, here, both components are synthesized. Thus, there is no 'real' signal to ensure some kind of stability. Since both signals are virtual, the ARL system is more likely to exhibit instabilities. There is no real data to `pull it back down to Earth' and these instabilities can grow limitlessly. Therefore, unless properly initialized, the system will not bootstrap. Furthermore, even when beginning to show some interaction, the system often locks up into some looping meaningless behaviour. Instabilities arise since both halves of the time series are completely synthesized with no real human tracking. Therefore, some modifications are being investigated for zero-user, two-computer configurations. However, this is not the most important mode of operation and remains a lower priority.

Thus we have enumerate several different modes of operation the ARL system can encompass. These include behaviour learning, interaction, simulation, prediction and filtering. The analysis of the ARL framework at a modular level allows such abstractions as well as the use of alternate modules and different data flow paths.