The Categorizability Assumption

From our experience developing controllers for autonomous robots, we observe that, in many realistic situations, the reward received by the robot depends only on a reduced subset of all the actions executed by the robot and that most of the sensor inputs are irrelevant to predict that reward. Thus, for example, the value resulting from the action of ``grasping the object in front of the robot" will depend on what the object is: the object the robot should bring to the user, an electrified cable, or an unimportant object. However, the result will probably be the same whether or not the robot is moving its cameras while grasping the object, if it is day or night, if the robot is, at the same time, checking the distance to the nearest wall, or if it can see a red light nearby or not (aspects, all of them, that may become important in other circumstances).

If an agent observes and acts in an environment where a reduced fraction of the available inputs and actuators have to be considered at a time, we say that the agent is in a categorizable environment.

Categorizability is not a binary predicate but a graded property. In the completely categorizable case, it would be necessary to pay attention to only one sensor/motor in each situation. On the other extreme of the spectrum, if all motors have to be carefully coordinated to achieve the task and the effect of each action could only be predicted by taking into account the value of all feature detectors, we would say that the environment is not categorizable at all.

Since robots have large collection of sensors providing a heterogeneous collection of inputs and many actuators affecting quite different degrees of freedom, our hypothesis is that, in robotic problems, environments are highly categorizable and, in those cases, an algorithm biased by the categorizability assumption would result advantageous.