Selecting an appropriate planning horizon for empowerment-based control

Determine the appropriate planning horizon for MPC-based empowerment and empowerment-based control more broadly, including how far into the future an agent must plan to produce stable and intelligent behavior, despite the non-monotonic relationship between planning horizon and performance.

Background

The proposed MPC-based empowerment controller is sensitive to hyperparameter choices, particularly the planning horizon. The paper notes that increasing the horizon does not necessarily improve performance; instead, the relationship between horizon length and performance is non-monotonic, which complicates both practical tuning and theoretical analysis.

The unresolved issue is therefore to identify how planning horizons should be selected for empowerment-based controllers and to characterize the future-planning range needed to produce stable and intelligent behavior across control settings. The authors frame this as an open challenge in their setting and in empowerment-based control more broadly.