""" Inference Script — Airport Flight Scheduling Environment ========================================================= Baseline agent using OpenAI-compatible client. Required env vars: API_BASE_URL — LLM endpoint (e.g. https://router.huggingface.co/v1) MODEL_NAME — model identifier HF_TOKEN — API key / HF token """ from __future__ import annotations import os import json import time import textwrap from typing import Any, Dict, List from openai import OpenAI # ── Config ───────────────────────────────────────────────────────────────────── API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1") API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY", "hf_placeholder") MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct") ENV_BASE_URL = os.getenv("ENV_BASE_URL", "http://localhost:7860") MAX_STEPS = 20 TEMPERATURE = 0.2 MAX_TOKENS = 400 # ── Env HTTP client ──────────────────────────────────────────────────────────── def env_post(path: str, payload: Dict = None) -> Dict: import urllib.request url = f"{ENV_BASE_URL}{path}" data = json.dumps(payload or {}).encode() req = urllib.request.Request(url, data=data, headers={"Content-Type": "application/json"}) with urllib.request.urlopen(req, timeout=30) as resp: return json.loads(resp.read()) def env_get(path: str) -> Dict: import urllib.request url = f"{ENV_BASE_URL}{path}" with urllib.request.urlopen(url, timeout=30) as resp: return json.loads(resp.read()) # ── Prompt builders ──────────────────────────────────────────────────────────── SYSTEM_PROMPT = textwrap.dedent(""" You are an AI Air Traffic Controller managing an airport. Each step you must output EXACTLY one JSON action object. Available action types: - assign_runway : {"action_type":"assign_runway","flight_id":"AA101","runway_id":"R1"} - assign_gate : {"action_type":"assign_gate","flight_id":"AA101","gate_id":"G1"} - delay_flight : {"action_type":"delay_flight","flight_id":"AA101","delay_minutes":5} - divert_flight : {"action_type":"divert_flight","flight_id":"AA101"} - clear_runway : {"action_type":"clear_runway","flight_id":"AA101","runway_id":"R1"} - declare_emergency: {"action_type":"declare_emergency","flight_id":"AA101"} Rules: 1. Never assign a flight to an OCCUPIED or CLOSED runway. 2. Respect wake turbulence separation (heavy before light = 5 min gap). 3. Match gate type: international flights → international gates (I*), domestic → domestic (D* or G*). 4. Declare emergency for any flight with EMERGENCY status immediately. 5. Output ONLY valid JSON. No explanation. """).strip() def obs_to_prompt(obs: Dict, step: int) -> str: clock = obs.get("clock", 0) events = obs.get("recent_events", []) flights = obs.get("flights", []) runways = obs.get("runways", []) gates = obs.get("gates", []) weather = obs.get("weather", {}) flight_lines = [] for f in flights: if f["status"] in ("landed", "diverted", "departed"): continue flight_lines.append( f" {f['flight_id']} | {f['weight_class']:6} | {f['status']:10} | " f"ETA:{f['eta_minutes']:.0f} | fuel:{f['fuel_remaining_min']:.0f} | " f"intl:{f['is_international']} | rwy:{f.get('assigned_runway','—')} | gate:{f.get('assigned_gate','—')}" ) rwy_lines = [ f" {r['runway_id']} ({r['name']}) — {r['status']} | avail@{r['available_at']:.0f}" for r in runways ] gate_lines = [ f" {g['gate_id']} | {'INTL' if g['is_international'] else 'DOM '} | {'BUSY' if g['occupied'] else 'FREE'}" for g in gates ] return textwrap.dedent(f""" Step: {step} | Clock: {clock:.0f} min Weather: {weather.get('condition','?')} | vis:{weather.get('visibility_km','?')}km | wind:{weather.get('wind_speed_kts','?')}kts Events: {'; '.join(events) if events else 'none'} ACTIVE FLIGHTS: {chr(10).join(flight_lines) if flight_lines else ' (none)'} RUNWAYS: {chr(10).join(rwy_lines)} GATES: {chr(10).join(gate_lines)} Output ONE JSON action:""").strip() # ── Main ─────────────────────────────────────────────────────────────────────── def run_task(client: OpenAI, task_id: str) -> float: print(f"\n{'='*60}") print(f"Task: {task_id}") print(f"{'='*60}") obs_data = env_post("/reset", {"task_id": task_id, "seed": 42}) print(f" Flights: {len(obs_data.get('flights', []))}") for step in range(1, MAX_STEPS + 1): if obs_data.get("episode_done") or obs_data.get("done"): break prompt = obs_to_prompt(obs_data, step) try: completion = client.chat.completions.create( model=MODEL_NAME, messages=[ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": prompt}, ], temperature=TEMPERATURE, max_tokens=MAX_TOKENS, ) raw = completion.choices[0].message.content or "{}" # Strip markdown fences if present raw = raw.strip().lstrip("```json").lstrip("```").rstrip("```").strip() action = json.loads(raw) except Exception as exc: print(f" Step {step}: LLM error ({exc}), using noop delay") # Find first active flight as fallback active = [f for f in obs_data.get("flights", []) if f["status"] not in ("landed","diverted","departed")] if active: action = {"action_type": "delay_flight", "flight_id": active[0]["flight_id"], "delay_minutes": 2} else: break try: result = env_post("/step", {"action": action, "task_id": task_id}) except Exception as exc: print(f" Step {step}: env error ({exc})") break obs_data = result.get("observation", obs_data) reward = result.get("reward", {}) done = result.get("done", False) print(f" Step {step:2d}: {action.get('action_type','?'):20s} | " f"reward: {reward.get('value', 0):+.2f} | " f"cumulative: {obs_data.get('score_so_far', 0):+.2f}") if done: print(f" Episode done at step {step}") break grade_data = env_get(f"/grade?task_id={task_id}") score = grade_data.get("score", 0.0) print(f" Final grade: {score:.4f}") return score def main(): client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY) tasks = [ "single_runway_landing", "multi_runway_gate_assignment", "storm_disruption_recovery", ] results = {} for task_id in tasks: try: score = run_task(client, task_id) results[task_id] = score except Exception as exc: print(f" ERROR on {task_id}: {exc}") results[task_id] = 0.0 time.sleep(1) print(f"\n{'='*60}") print("BASELINE RESULTS") print(f"{'='*60}") for task_id, score in results.items(): print(f" {task_id:<40} {score:.4f}") avg = sum(results.values()) / len(results) print(f" {'AVERAGE':<40} {avg:.4f}") if __name__ == "__main__": main()