"""Bes modeli de fit eder. Nonlineerlik sadece C'de; C verildiginde model lineer. Bu yuzden profil edilmis en kucuk kareler kullaniyoruz (scipy gerekmiyor).""" import os # Paths are resolved relative to this script so the package is self-contained. _HERE = os.path.dirname(os.path.abspath(__file__)) DATA = os.path.normpath(os.path.join(_HERE, "..", "..", "data")) import csv, numpy as np, collections # ---- veri ---- lat = {} for r in csv.DictReader(open(os.path.join(DATA, "latency_210.csv"), encoding="utf-8-sig")): lat[(int(r["qIdx"]), int(float(r["B_Mbps"])), int(r["trial"]))] = float(r["latency_ms"]) bw = {} for r in csv.DictReader(open(os.path.join(DATA, "latency_trials_raw_210.csv"), encoding="utf-8")): bw[(int(r["qIdx"]), int(r["target_Mbps"]), int(r["trial"]))] = float(r["achieved_Mbps"]) keys = sorted(lat) assert len(keys) == 210 and all(k in bw for k in keys) q = np.array([k[0] for k in keys]) tgt = np.array([k[1] for k in keys]) tri = np.array([k[2] for k in keys]) L = np.array([lat[k] for k in keys]) b = np.array([bw[k] for k in keys]) train = np.isin(tri, [1, 2, 3]) test = np.isin(tri, [4, 5]) def lstsq(X, y): beta, *_ = np.linalg.lstsq(X, y, rcond=None) return beta def metrics(y, yh): e = y - yh return (np.mean(np.abs(e)), np.sqrt(np.mean(e ** 2)), 1 - np.sum(e ** 2) / np.sum((y - y.mean()) ** 2)) # ---------- tasarim matrisleri (C verildiginde lineer) ---------- def X_full(bb, qq, C): # L0 + beta*b + theta/(C-b) + gamma_q cols = [np.ones_like(bb), bb, 1.0 / (C - bb)] for j in range(2, 8): cols.append((qq == j).astype(float)) return np.column_stack(cols) def X_noloc(bb, qq, C): # L0 + beta*b + theta/(C-b) return np.column_stack([np.ones_like(bb), bb, 1.0 / (C - bb)]) def X_mm1(bb, qq, C): # L0 + theta/(C-b) return np.column_stack([np.ones_like(bb), 1.0 / (C - bb)]) def X_poly(bb, qq, C=None): # poly22 return np.column_stack([np.ones_like(bb), bb, qq, bb ** 2, bb * qq, qq ** 2]) def X_persat(bb, qq, C=None, Bth=200.0): # L0(q)+alpha(q)ln(1+b)+beta(q)max(0,b-Bth) cols = [] f1, f2 = np.log1p(bb), np.maximum(0.0, bb - Bth) for j in range(1, 8): m = (qq == j).astype(float) cols += [m, m * f1, m * f2] return np.column_stack(cols) def fit_profiled(Xfun, bb, qq, yy): """C uzerinde 1-B arama + her C icin lineer cozum.""" lo = bb.max() + 1e-3 best = None for C in np.concatenate([np.linspace(lo, lo + 50, 400), np.linspace(lo + 50, 3000, 3000)]): X = Xfun(bb, qq, C) p = lstsq(X, yy) r = yy - X @ p s = float(r @ r) if best is None or s < best[0]: best = (s, C, p) # yerel incelt s, C, p = best step = 1.0 for _ in range(60): improved = False for cand in (C - step, C + step): if cand <= lo: continue X = Xfun(bb, qq, cand) pp = lstsq(X, yy) r = yy - X @ pp ss = float(r @ r) if ss < s: s, C, p, improved = ss, cand, pp, True if not improved: step /= 2 if step < 1e-6: break return C, p MODELS = { "(a) proposed full": (X_full, True), "(b) no-location": (X_noloc, True), "(c) minimal M/M/1-type": (X_mm1, True), "(d) polynomial poly22": (X_poly, False), "(e) per-location sat.": (X_persat, False), } print("=" * 78) print("1) TAM VERI SETI ILE (a) MODELI -- makaledeki Tablo 3 ile karsilastirma") print("=" * 78) C, p = fit_profiled(X_full, b, q, L) X = X_full(b, q, C) mae, rmse, r2 = metrics(L, X @ p) print(f" L0 = {p[0]:8.3f} ms (makale 8.639)") print(f" beta = {p[1]:8.4f} ms/Mbps (makale 0.0305)") print(f" theta = {p[2]:8.3f} ms*Mbps (makale 486.885)") print(f" C = {C:8.3f} Mbps (makale 611.395)") gp = [-0.470, -1.449, -1.306, 0.782, -0.476, -1.681] for i, j in enumerate(range(2, 8)): print(f" gamma{j} = {p[3 + i]:8.3f} ms (makale {gp[i]:+.3f})") print(f" MAE={mae:.3f} RMSE={rmse:.3f} R2={r2:.4f} (makale 2.52 / 3.13 / 0.842)") print() print("=" * 78) print("2) TRIAL-BASED TRAIN(1-3)/TEST(4-5) -- makaledeki Tablo 4 ile karsilastirma") print("=" * 78) ref = {"(a) proposed full": (2.654, 3.269, 0.835), "(b) no-location": (2.655, 3.329, 0.829), "(c) minimal M/M/1-type": (2.892, 3.633, 0.797), "(d) polynomial poly22": (2.724, 3.382, 0.824), "(e) per-location sat.": (1.787, 2.322, 0.917)} coef = {} for name, (Xf, hasC) in MODELS.items(): if hasC: Ci, p = fit_profiled(Xf, b[train], q[train], L[train]) else: Ci = None p = lstsq(Xf(b[train], q[train]), L[train]) yh = (Xf(b[test], q[test], Ci) if hasC else Xf(b[test], q[test])) @ p m = metrics(L[test], yh) coef[name] = (Ci, p) r = ref[name] print(f" {name:24s} test MAE={m[0]:6.3f} RMSE={m[1]:6.3f} R2={m[2]:6.3f}" f" | makale {r[0]:.3f}/{r[1]:.3f}/{r[2]:.3f}") print() print("=" * 78) print("3) EGITIM SETI KATSAYILARI (Appendix A icin)") print("=" * 78) Ci, p = coef["(a) proposed full"] print("(a) proposed full model") print(f" L0={p[0]:.3f} beta={p[1]:.4f} theta={p[2]:.2f} C={Ci:.2f}") print(" " + " ".join(f"g{j}={p[3+i]:+.3f}" for i, j in enumerate(range(2, 8)))) Ci, p = coef["(b) no-location"] print(f"(b) no-location ablation\n L0={p[0]:.3f} beta={p[1]:.4f} theta={p[2]:.2f} C={Ci:.2f}") Ci, p = coef["(c) minimal M/M/1-type"] print(f"(c) minimal M/M/1-type\n L0={p[0]:.3f} theta={p[2-1]:.2f} C={Ci:.2f}") _, p = coef["(d) polynomial poly22"] print("(d) polynomial poly22") print(" " + " ".join(f"a{i}={v:+.5g}" for i, v in enumerate(p))) _, p = coef["(e) per-location sat."] print("(e) per-location saturation (Bth=200 Mbps)") print(" q : L0(q) alpha(q) beta(q)") for j in range(7): print(f" Q{j+1}: {p[3*j]:9.3f} {p[3*j+1]:10.3f} {p[3*j+2]:9.4f}")