Why Complex Intelligent Decisions Must Be Decomposed (and How)

1. Five Dimensions of Complexity

State‑space explosion : high‑dimensional variables, mix of continuous and discrete. Consequence : dimensional disaster makes computation infeasible.

Multi‑objective conflict : efficiency vs safety, short‑term vs long‑term, local vs global. Consequence : objective function cannot be defined, Pareto frontier unsolvable.

Multiple time scales : strategic (year/month) → tactical (day/hour) → execution (second/millisecond). Consequence : a single algorithm cannot satisfy both responsiveness and global optimality.

Heterogeneous uncertainty : perception noise, model error, opponent game, environmental change. Consequence : a single robustness assumption becomes either too strong or too weak.

Multi‑agent interaction : collaboration, competition, limited communication. Consequence : centralized solving hits communication and compute bottlenecks.

2. Four Classic Decomposition Paradigms

Hierarchical Decomposition

Splits decisions by abstraction layers.

Strategic layer (Deliberative) : "Where to go" – goal selection, route planning.

Tactical layer (Tactical) : "How to go" – behavior sequencing, resource scheduling.

Execution layer (Reactive) : "Avoid obstacles" – reflex control, real‑time response.

Autonomous‑driving example

Upper layer (navigation): from Beijing to Shanghai, which highway, estimated travel time – second‑to‑minute decisions.

Middle layer (behavior): congestion ahead, lane‑change decision; yellow light at intersection – ~100 ms decisions.

Lower layer (trajectory): steering wheel angle, throttle depth – ~10 ms decisions.

Using a single optimizer would require a state space that combines map topology, vehicle dynamics, traffic participants, and time, rendering the problem unsolvable.

Functional Modularization

Decisions are orthogonally divided by function, allowing independent iteration and verification.

Perception decision : "What the world is" – state estimation.

Prediction decision : "What the world will be" – intent inference of other agents.

Planning decision : "What to do" – ego‑vehicle behavior.

Verification decision : "Is it safe" – safety monitoring and guard mechanisms.

Safety‑verification layer can act as an independent monitor; if the planning layer fails, the monitor can trigger protective actions.

Time‑Scale Decomposition

Different time scales employ different decision methodologies.

Millisecond level : reflex control – pre‑computed lookup tables, MPC hard constraints.

Second level : short‑term optimization – receding horizon control (RHC), reinforcement learning.

Minute level : task planning – symbolic AI, constraint solving, heuristic search.

Hour‑plus level : strategic decision – game theory, multi‑agent negotiation, human intervention.

Upper layers set abstract goals or constraints; lower layers provide feasibility feedback that constrains the upper layers.

Spatial/Responsibility‑Domain Decomposition

Geographic partition : different regions handled by distinct decision nodes (e.g., air‑traffic control sectors).

Functional partition : perception, decision, execution distributed across physical nodes (edge‑cloud collaboration).

Multi‑agent partition : each agent makes autonomous decisions and coordinates via negotiation or game‑theoretic mechanisms.

3. Core Challenge After Decomposition: Interface Design

1. Information Compression

// ❌ Wrong: "At second 5, turn steering wheel 15°"
// ✅ Correct: "Stay centered in lane, keep distance to front vehicle ≥ 2 seconds"

The lower layer decides how to achieve the constraint.

2. Consistency Guarantees

Forward consistency : lower‑level plans must satisfy upper‑level constraints (feasibility verification).

Backward consistency : upper‑level plans must consider lower‑level capability boundaries (reachability feedback).

3. Exception Fallback Mechanisms

Normal operation : three‑layer cooperation.

Tactical layer failure : execution layer runs in safety mode (emergency braking, pull‑over).

Execution layer failure : hardware‑level protection (power cut, mechanical brake).

4. Conceptual Distinctions

Divide‑and‑conquer algorithm : recursively breaks a problem into atomic sub‑problems. Relation : decision decomposition emphasizes semantic/functional boundaries rather than pure mathematical split.

Federated learning : data stays local, models move. Relation : a form of spatial decomposition implemented in a distributed manner.

End‑to‑end learning : bypasses explicit intermediate representations. Relation : appears opposed to explicit decomposition, but large models implicitly achieve a "soft" split via attention mechanisms.

Micro‑service architecture : software‑engineered modularity. Relation : maps decision decomposition onto engineering implementation.

5. When Not to Decompose

Strongly coupled systems : inter‑subproblem coupling is so high that interface complexity exceeds the original problem.

Data‑driven tasks with fuzzy boundaries : implicit decisions in perception may be better served by end‑to‑end approaches.

Small‑scale global‑optimal problems : when problem size is manageable, centralized solving can be simpler and more efficient.

Core Takeaway

Complex intelligent decisions must be decomposed because cognitive and computational resources are limited and the problem structure must be matchable to manageable sub‑problems. Effective decomposition follows the principle of “high cohesion, low coupling”: each decision unit is internally cohesive, and units interact through minimal, clear interfaces, turning unmanageable complexity into manageable complexity while preserving robustness and optimality.