Contributed talk in Swarm Behaviour 1, Aug. 1, 2019, 12:30 p.m. in room NUBS 2.03

On information-optimal scripting of actions

Bente Riegler, Daniel Polani

watch Publication Animals and humans encounter many tasks which permit ritualized behaviours, essentially fixed action sequences or ``scripts'', similar to options known from Reinforcement Learning, but proceeding without intermediate decisions. While running a script, they proceed in an open-loop fashion. However, when they are completed, an agent needs to decide whether to perform a basic action or to trigger a script regarding the particular task. Here we study when including such action scripts (i.e. behaviour rituals) is advantageous from the point of view of the relevant information required to take the decision to start such a sequence, also depending on the tasks. To achieve this, we modify the relevant information framework to include sequences of basic actions in the overall selection of possible actions.