Sensor feedback

An interesting I-map is obtained by removing all but the last sensor observation from the history I-state. This yields an I-map, $ {{\kappa}_{sf}}
: {\cal I}_{hist}\rightarrow Y$, which is defined as $ {{\kappa}_{sf}}({\eta}_k) = y_k$. The model is referred to as sensor feedback. In this case, all decisions are made directly in terms of the current sensor observation. The derived I-space is $ Y$, and a plan on the derived I-space is $ \pi: Y \rightarrow U$, which is called a sensor-feedback plan. In some literature, this may be referred to as a purely reactive plan. Many problems for which solutions exist in the history I-space cannot be solved using sensor feedback. Neglecting history prevents the complicated deductions that are often needed regarding the state. In some sense, sensor feedback causes short-sightedness that could unavoidably lead to repeating the same mistakes indefinitely. However, it may be worth determining whether such a sensor-feedback solution plan exists for some particular problem. Such plans tend to be simpler to implement in practice because the actions can be connected directly to the sensor output. Certainly, if a sensor-feedback solution plan exists for a problem, and feasibility is the only concern, then it is pointless to design and implement a plan in terms of the history I-space or some larger derived I-space. Note that this I-map is sufficient, even though it ignores the entire history.

Steven M LaValle 2012-04-20