CONCEPT Cited by 1 source
Hierarchical policy decomposition¶
Definition¶
Hierarchical policy decomposition is the principle of splitting a complex decision system into two or more layers operating at different time horizons and abstraction levels: a slow/strategic layer that sets long-term goals and constraints, and a fast/tactical layer that makes immediate decisions within those bounds.
The pattern appears across domains: - Robotics / autonomous driving: slower planning layer (goals, constraints, long horizons) + faster control loops (steering, braking) - LLM agents: deliberate planning + rapid step-by-step tool use - Notification systems: weekly frequency planning + per-opportunity message selection
The key insight is that decoupling time horizons prevents short-term optimization from undermining long-term objectives — e.g., maximizing today's click-through without accounting for cumulative fatigue.
Seen in¶
- sources/2026-06-19-netflix-thinking-fast-slow-for-a-personalized-notification-system — Netflix's Slow/Fast notification architecture: Slow Policy sets personalized weekly pacing, Fast Policy selects messages in real time. Produced their largest production metric lift.