LLMPC: LLM 기반 Model Predictive Control 플래닝 프레임워크

import openai
import json

def evaluate_plan(plan: str, constraints: dict) -> tuple[float, list[str]]:
    """플랜이 제약 조건을 얼마나 충족하는지 평가. 위반 항목 리스트 반환"""
    violations = []
    # 여기에 도메인별 제약 검사 로직 구현
    cost = len(violations)  # 위반 수를 비용으로 사용
    return cost, violations

def llmpc_plan(
    task: str,
    constraints: dict,
    system_prompt: str,
    max_iterations: int = 7,
    plans_per_iter: int = 3,
    model: str = "gpt-4o"
) -> str:
    client = openai.OpenAI()
    best_plan = ""
    best_cost = float("inf")
    feedback_string = ""

    for step in range(1, max_iterations + 1):
        # 현재 상태 + 피드백을 넣어 K개 후보 요청
        instruction = f"""
STEP {step}/{max_iterations}
TASK: {task}

Your current best plan is:
{best_plan if best_plan else 'No plan yet.'}

{feedback_string}

Propose {plans_per_iter} different plans separated by '---'.
"""
        response = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": instruction}
            ]
        )
        
        raw = response.choices[0].message.content
        candidates = raw.split("---")
        
        # 각 후보를 평가해서 최선 선택 (MPC의 비용 함수 선택 단계)
        for candidate in candidates:
            candidate = candidate.strip()
            if not candidate:
                continue
            cost, violations = evaluate_plan(candidate, constraints)
            if cost < best_cost:
                best_cost = cost
                best_plan = candidate
                feedback_string = "Unmet constraints:\n" + "\n".join(f"- {v}" for v in violations) if violations else ""
        
        if best_cost == 0:
            print(f"완벽한 플랜을 {step}번째 반복에서 찾음!")
            break
    
    return best_plan

# 사용 예시
task = "Visit 5 cities in 10 days with given flight constraints..."
constraints = {"days": 10, "cities": ["Paris", "London", "Rome", "Berlin", "Madrid"]}
system_prompt = "You are an expert travel planner. Propose valid trip plans."
result = llmpc_plan(task, constraints, system_prompt, max_iterations=7, plans_per_iter=3)