ralstp-consultant（拉尔斯特普顾问）

RALSTP Consultant

Based on "Recursive Agents and Landmarks Strategic-Tactical Planning (RALSTP)" by Dorian Buksz, King's College London, 2024.

Core Concepts (from the thesis)

1. Agents Identification

Definition: Agents are objects with dynamic types that are active during goal state search.

How to identify:

• Dynamic type = appears as first argument of a predicate in any action's effects
• Static type = never appears in action effects
• Example: In Driverlog, truck and driver are dynamic (they're in drive action effects), but location is static

Real PDDL Example (RTAM Domain):

(:types  
   ambulance police_car tow_truck fire_brigade - vehicle
   acc_victim vehicle car - subject
   ...
)

• Agents: ambulance, police_car, tow_truck, fire_brigade (appear in action effects like at, available, busy)
• Passive: acc_victim, car (acted upon but don't act)

2. Passive Objects

Objects that are NOT agents — things being acted upon but don't act themselves.

• Packages, cargo, data, files, victims in RTAM

3. Agent Dependencies

Definition: Relationships between agents based on what preconditions they satisfy for other agents.

Types:

• Independent — agents that don't depend on each other
• Dependent — agents that need other agents' preconditions satisfied
• Conflicting — agents that interfere with each other

4. Entanglement

Definition: When agents fight for shared resources (time, space, locations, etc.)

Measurement:

• Count of shared predicates
• Conflict frequency in goal states

Real PDDL Example (RTAM - Road Traffic Accident):

(:durative-action confirm_accident
   :parameters (?V - police_car ?P - subject ?A - accident_location)
   :condition (and (at start (at ?V ?A)) (at start (at ?P ?A)) ...)
   :effect (and (at end (certified ?P)) ...)
)

(:durative-action untrap
   :parameters (?V - fire_brigade ?P - acc_victim ?A - accident_location)
   :condition (and (at start (certified ?P)) (at start (available ?V)) ...)
)

• Entanglement: police_car must certify BEFORE fire_brigade can untrap
• Resource conflict: Both need to be at same accident_location
• Availability: fire_brigade busy during untrap → others must wait

5. Landmarks

Definition: Facts that must be true in any valid plan (from goals back to initial state).

Types:

• Fact landmarks — propositions that must hold
• Action landmarks — actions that must be executed
• Relaxed landmarks — landmarks considering only positive effects (ignoring deletes)

Real PDDL Example (RTAM - sequential dependencies):

Goal: (delivered victim1) ∧ (delivered car1)

Required sequence of fact landmarks:
1. (certified victim1)     ← police must confirm
2. (untrapped victim1)     ← fire must free them
3. (aided victim1)         ← ambulance must treat
4. (loaded victim1 ambulance) ← ambulance must load
5. (at victim1 hospital)   ← deliver to hospital
6. (delivered victim1)     ← FINAL

Action landmarks:
- confirm_accident → untrap → first_aid → load_victim → unload_victim → deliver_victim

6. Strategic vs Tactical

• Strategic: Abstract planning level. Solve "what needs to happen first" ignoring details.
• Tactical: Detailed execution level. Solve "exactly how to do it".

7. Difficulty Metrics

From the thesis, difficulty increases with:

• More agents in goal state
• More entangled agents (conflicting dependencies)
• More inactive dynamic objects not in goal

Buksz Complexity Score ≈ Agent Count × Entanglement Factor

Implementation Note (Natural Language vs PDDL)

This skill operates in two modes:

1. Conceptual Mode (Default): Uses the LLM to apply RALSTP methodology to natural language problems (e.g., "Plan a marketing launch"). No PDDL files are required. The agent identifies Agents/Landmarks conceptually.
2. Formal Mode (Optional): If you provide PDDL domain/problem files, the included scripts/analyze.py can be run to mathematically extract agents and landmarks.

*The instructions below apply to both modes, but "Real PDDL Examples" are provided for technical context.*

Usage

For any complex problem, just describe it and I'll apply RALSTP:

RALSTP analyze: I need to migrate 1000 VMs from datacentre A to B with minimal downtime

Output Format

## RALSTP Analysis

### Agents Identified
- [list agents and their types]

### Passive Objects  
- [list objects being acted upon]

### Dependency Graph
- [which agents depend on which]

### Difficulty Assessment
- Agent Count: X
- Entanglement: Low/Medium/High
- Estimated Complexity: [score]

### Strategic Phase
- [high-level plan ignoring details]

### Tactical Phase
- [detailed execution]

### Decomposition Suggestion
- Split by: [agent type / landmark / location]
- Parallelize: [what can run concurrently]
- Risks: [potential conflicts/entanglements]

When to Use

USE for:

• Multi-step workflows with multiple actors
• Migration/tasks with dependencies
• Resource contention problems
• Complex orchestrations

SKIP for:

• Simple Q&A
• Single-task problems

Reference

PhD Thesis: "Recursive Agents and Landmarks Strategic-Tactical Planning (RALSTP)" — Dorian Buksz, King's College London, 2024.

Example: RTAM Domain (IPC-2014)

Domain: Road Traffic Accident Management

Source: https://github.com/potassco/pddl-instances/tree/master/ipc-2014/domains/road-traffic-accident-management-temporal-satisficing

Full Analysis

Agents (4):

• ambulance — transports victims to hospital
• police_car — certifies accident/victims
• tow_truck — recovers vehicles
• fire_brigade — untraps victims, extinguishes fires

Passive Objects:

• acc_victim — people needing help
• car — vehicles involved in accident
• accident_location, hospital, garage

Dependencies (Critical Path):

police_car → fire_brigade → ambulance → hospital
     ↓            ↓           ↓
  certify      untrap       deliver

Landmarks Chain (must execute in order):

1. confirm_accident (police at scene)
2. untrap (fire frees victim)
3. first_aid (ambulance treats)
4. load_victim → unload_victim → deliver_victim
5. load_car → unload_car → deliver_vehicle

Entanglement:

• Multiple vehicles must be at same location (accident scene)
• Vehicles have limited availability (busy during actions)
• Sequence constraints: can't deliver before certify

Difficulty: High — 4 agents, tight dependencies, shared locations