Three policies, chosen by who proposed the goal.

Mirrors tbp/monty/frameworks/models/motor_policy_selectors.py::DistantPolicySelector, the selector a distant agent runs. A distant agent can be asked for three different kinds of step in one run, and the sender of the goal is what says which:

• a learning module proposed it (:gsg): the module wants a hypothesis tested, so the agent teleports — JumpToGoal — and spends the following step judging where it landed;
• a sensor module proposed it (:sm): the sensor wants to look at something, so the agent turns — LookAtGoal;
• nobody proposed anything: the default policy wanders.

A jump in flight takes priority over everything else, including a new learning-module goal, because the agent is half-way through an experiment it has to finish: JumpToGoal is asked first, and only if it has nothing to say does the step fall through to the other two. Of each kind, the most confident goal is the one pursued.

Why it holds policies rather than functions

SinglePolicySelector holds its policy as a five-argument function, because a step is all it needs. This one needs more: it reads the jump's status to know whether a jump is in flight, and it resets and reports on all three policies. So it holds the three policy structs and threads each one's state, which is Monty's shape too — its selector holds policy objects and calls their methods.