Learning Where To Go without Knowing Where That Is: The Acquisition of a Non-reactive Mobot Behaviour by Explicitation

In the path-imitation task, one agent traces out a path through a second agent's sensory eld. The second agent then has to reproduce that path exactly, i.e. move through the sequence of locations visited by the rst agent. This is a non-trivial behaviour whose acquisition might be expected to involve special-purpose (i.e., strongly biased) learning machinery. However, the present paper shows this is not the case. The behaviour can be acquired using a fairly primitive learning regime provided that the agent's environment can be made to pass through a speciic sequence of dynamic states.