"""Study 1: simulator-in-the-loop vs formula model. Part A (prediction): N random feasible plans; formula delay vs SUMO delay (SEEDS_A seeds each) -> correlation, scatter. Part B (optimization): derivative-free optimizer (Nelder-Mead) whose objective is mean SUMO delay (SEEDS_B seeds, common random numbers), same knobs as the model (C, g_main per light, offsets), warm-started from the formula optimum. Fixed budget of SUMO evaluations. Best plan saved as plans/study1-sumo-loop.json for the main comparison. Outputs: build/study1/{prediction.csv, optimization.csv, summary.json, prediction.png, convergence.png} Usage: study1.py [--n-random 30] [--budget 120] """ import sys, json, csv, time, pathlib, shutil, statistics import numpy as np from scipy.optimize import minimize from concurrent.futures import ThreadPoolExecutor sys.path.insert(0, str(pathlib.Path(__file__).parent)) from common import BUILD, corridor import plans as P, plan_to_sumo, run_sumo, formula OUT = BUILD/"study1" SEEDS_A, SEEDS_B = (1, 2), (1, 2) def plan_from(cor, C, g, offs, name, source): ids = [l["id"] for l in cor["lights"]] return dict(name=name, source=source, C=float(C), L_s=cor["params"]["L_s"], lights={i: dict(g_main=float(gi), offset_s=float(o % C)) for i, gi, o in zip(ids, g, offs)}, notes="") def feasible(cor, C, g): return min(formula.constraints(cor, C, g)) >= -1e-6 def sumo_delay(plan, seeds): """Mean SUMO delay over seeds; None if the plan can't be built or SUMO fails (logged).""" try: plan_to_sumo.write(plan) with ThreadPoolExecutor(max_workers=len(seeds)) as ex: res = list(ex.map(lambda s: run_sumo.run_one(plan["name"], s), seeds)) except Exception as e: print(f" ! {plan['name']} failed: {str(e)[:120]}") shutil.rmtree(run_sumo.RUNS/plan["name"], ignore_errors=True) return None shutil.rmtree(run_sumo.RUNS/plan["name"], ignore_errors=True) (plan_to_sumo.SUMO/"plans"/f"{plan['name']}.tls.add.xml").unlink(missing_ok=True) return statistics.mean(r["delay_veh_h_per_h"] for r in res) def random_plan(cor, rng): Pm = cor["params"]; n = len(cor["lights"]) for _ in range(5000): C = rng.uniform(Pm["C_min_s"], Pm["C_max_s"]); g = [] for l in cor["lights"]: A = l["approaches"]; L = formula.light_L(cor, l) lo = max(Pm["g_min_s"], max(A[d]["q_vph"]*C/(Pm["x_max"]*A[d]["s_vph"]) for d in ("NB", "SB"))) hi = C-L-max(Pm["g_min_s"], max(A[d]["q_vph"]*C/(Pm["x_max"]*A[d]["s_vph"]) for d in ("EB", "WB"))) if hi <= lo: break g.append(rng.uniform(lo, hi)) if len(g) == n and feasible(cor, C, g): return C, g, [0.0]+list(rng.uniform(0, C, n-1)) raise RuntimeError("no feasible random plan found") def part_a(cor, n): rng = np.random.default_rng(7); rows = [] for k in range(n): C, g, o = random_plan(cor, rng) p = plan_from(cor, C, g, o, f"_s1a_{k}", "study1 random") fd = formula.evaluate(cor, C, g, o) sd = sumo_delay(p, SEEDS_A) if sd is None: continue rows.append(dict(k=k, C=C, formula=fd, sumo=sd)); print(f" A {k+1}/{n}: C={C:5.1f} formula {fd:7.1f} sumo {sd:7.1f}") return rows def part_b(cor, budget): ids = [l["id"] for l in cor["lights"]]; n = len(ids); Pm = cor["params"] start = P.load("formula-optimum-standin") z0 = [start["C"]]+[start["lights"][i]["g_main"] for i in ids]+[start["lights"][i]["offset_s"]/start["C"] for i in ids[1:]] hist = []; t0 = time.time() def obj(z): C = float(np.clip(z[0], Pm["C_min_s"], Pm["C_max_s"])); g = list(z[1:1+n]); w = [0.0]+[float(x) % 1 for x in z[1+n:]] if not feasible(cor, C, g): pen = -min(formula.constraints(cor, C, g)) val = 1e4+pen else: val = sumo_delay(plan_from(cor, C, g, [x*C for x in w], f"_s1b_{len(hist)}", "study1 eval"), SEEDS_B) val = 1e4 if val is None else val hist.append(dict(eval=len(hist)+1, value=val, C=C, g=g, w=w, t=time.time()-t0)) best = min(h["value"] for h in hist) print(f" B eval {len(hist):3d}: {val:8.1f} best {best:8.1f}") return val simplex = [np.array(z0, float)] steps = [8.0]+[4.0]*n+[0.08]*(n-1) for k, s in enumerate(steps): v = np.array(z0, float); v[k] += s; simplex.append(v) minimize(obj, z0, method="Nelder-Mead", options=dict(maxfev=budget, initial_simplex=np.array(simplex), xatol=1e-2, fatol=0.5)) b = min(hist, key=lambda h: h["value"]) plan = plan_from(cor, b["C"], b["g"], [x*b["C"] for x in b["w"]], "study1-sumo-loop", f"Study 1: Nelder-Mead on SUMO delay ({len(hist)} evaluations x {len(SEEDS_B)} seeds), warm start = formula optimum") P.save(plan) return hist, plan def figures(rows, hist): import matplotlib; matplotlib.use("Agg"); import matplotlib.pyplot as plt if rows: f, ax = plt.subplots(figsize=(5, 4)) ax.scatter([r["formula"] for r in rows], [r["sumo"] for r in rows], s=18) ax.set_xlabel("formula model delay (veh-h/h)"); ax.set_ylabel("SUMO delay (veh-h/h)") ax.set_title("Study 1A: does the formula rank plans like SUMO?"); f.tight_layout(); f.savefig(OUT/"prediction.png", dpi=150) if hist: f, ax = plt.subplots(figsize=(5, 3.5)) vals = [h["value"] for h in hist if h["value"] < 1e4]; best = np.minimum.accumulate([h["value"] for h in hist]) ax.plot(range(1, len(best)+1), best); ax.set_xlabel("SUMO evaluations"); ax.set_ylabel("best delay so far (veh-h/h)") ax.set_title("Study 1B: simulator-in-the-loop optimization"); f.tight_layout(); f.savefig(OUT/"convergence.png", dpi=150) def main(): args = sys.argv[1:] n_random = int(args[args.index("--n-random")+1]) if "--n-random" in args else 30 budget = int(args[args.index("--budget")+1]) if "--budget" in args else 120 OUT.mkdir(parents=True, exist_ok=True); cor = corridor() t = time.time(); rows = part_a(cor, n_random); ta = time.time()-t from scipy.stats import spearmanr, pearsonr fin = [r for r in rows if r["formula"] < 900] rho = spearmanr([r["formula"] for r in fin], [r["sumo"] for r in fin]).statistic if len(fin) > 2 else None r_p = pearsonr([r["formula"] for r in fin], [r["sumo"] for r in fin]).statistic if len(fin) > 2 else None with open(OUT/"prediction.csv", "w", newline="") as f: w = csv.DictWriter(f, fieldnames=list(rows[0])); w.writeheader(); w.writerows(rows) t = time.time(); hist, plan = part_b(cor, budget); tb = time.time()-t with open(OUT/"optimization.csv", "w", newline="") as f: w = csv.writer(f); w.writerow(["eval", "sumo_delay", "C", "seconds"]) for h in hist: w.writerow([h["eval"], round(h["value"], 3), round(h["C"], 2), round(h["t"], 1)]) t = time.time(); r_formula = formula.optimize(cor, starts=30); tf = time.time()-t summary = dict(prediction=dict(n_plans=len(rows), n_finite=len(fin), spearman=rho, pearson=r_p, seconds=round(ta)), sumo_loop=dict(evaluations=len(hist), seeds_per_eval=len(SEEDS_B), best_delay_2seeds=min(h["value"] for h in hist), seconds=round(tb), plan="plans/study1-sumo-loop.json"), formula_optimizer=dict(seconds=round(tf, 1), starts=30, formula_delay=r_formula["total_veh_h"])) json.dump(summary, open(OUT/"summary.json", "w"), indent=1, default=float) figures(rows, hist) print(json.dumps(summary, indent=1, default=float)) if __name__ == "__main__": main()