ccevolve/tx_workspace/ac1_ours/attack8_tile.py

"""Attack 8: Tile the best small-n solution to create larger sequences.
Tiling preserves local structure better than interpolation."""
import time
import numpy as np
from scipy.signal import fftconvolve
from scipy.optimize import minimize

def actual_obj(a):
    a = np.maximum(a, 0); n = len(a)
    conv = fftconvolve(a, a); s = np.sum(a)
    if s < 0.01: return float('inf')
    return 2.0 * n * np.max(conv) / s**2

def make_lp(n, p):
    def f(a):
        a = np.maximum(a, 1e-12); S = np.sum(a)
        conv = fftconvolve(a, a); conv = np.maximum(conv, 1e-30)
        lc = np.log(conv); lcm = np.max(lc); lcs = lc - lcm
        e = np.exp(p * lcs); se = np.sum(e)
        Lp = np.exp(lcm) * se ** (1.0/p)
        obj = 2.0 * n * Lp / S**2
        w = (se ** ((1-p)/p)) * np.exp((p-1) * lcs)
        G = fftconvolve(w, a[::-1], mode='valid')
        G = G[:n] if len(G) >= n else np.pad(G, (0, n-len(G)))
        return obj, 2.0 * n / S**2 * (2 * G - 2.0 * Lp / S)
    return f

# Load the best n=300 template  
a300 = np.load('/workspace/best_sequence.npy')
print(f"n=300 template: {actual_obj(a300):.7f}")

# Try different upscaling strategies
strategies = []

# 1. Tile (repeat)
for k in [4, 7, 10]:
    a_tiled = np.tile(a300, k)[:2000]
    strategies.append(("tile_"+str(k), a_tiled))

# 2. Tile with overlap blending
for k in [7, 10]:
    a_tiled = np.tile(a300, k + 1)
    # Blend at tile boundaries
    n_base = len(a300)
    overlap = n_base // 4
    for t in range(1, k):
        start = t * n_base - overlap
        for j in range(2 * overlap):
            w = j / (2 * overlap)
            idx = start + j
            if idx < len(a_tiled):
                a_tiled[idx] = (1-w) * a_tiled[idx] + w * a_tiled[idx]
    a_tiled = a_tiled[:2000]
    strategies.append(("tile_blend_"+str(k), a_tiled))

# 3. Zero-insertion upsampling
a_upsampled = np.zeros(2000)
ratio = 2000 / 300
for i in range(300):
    j = int(i * ratio)
    if j < 2000:
        a_upsampled[j] = a300[i]
# Smooth
from scipy.ndimage import uniform_filter1d
a_smooth = uniform_filter1d(a_upsampled, int(ratio) + 1)
a_smooth = np.maximum(a_smooth, 1e-10)
strategies.append(("zero_insert", a_smooth))

# 4. Standard interpolation (baseline)
a_interp = np.interp(np.linspace(0, 1, 2000), np.linspace(0, 1, 300), a300)
strategies.append(("interp", a_interp))

# Test each strategy
best_val = float('inf')
best_a = None
bounds = [(1e-10, 1000.0)] * 2000

for name, a0 in strategies:
    a0 = np.maximum(a0, 1e-10)
    init_obj = actual_obj(a0)
    
    t0 = time.time()
    for p in [32, 128, 512, 2048, 8192, 32768, 131072, 524288]:
        if time.time() - t0 > 30: break
        res = minimize(make_lp(2000, p), a0, method='L-BFGS-B', jac=True, bounds=bounds,
                      options={'maxiter': 10000, 'ftol': 1e-16, 'gtol': 1e-15})
        a0 = np.maximum(res.x, 1e-10)
    
    val = actual_obj(a0)
    print(f"{name:20s}: init={init_obj:.4f} → final={val:.7f} ({time.time()-t0:.0f}s)")
    
    if val < best_val:
        best_val = val
        best_a = a0.copy()

print(f"\nBest tiling: {best_val:.7f}")
print(f"Current best: {actual_obj(np.load('/workspace/best_n2000_ultrahigh.npy')):.7f}")

if best_val < actual_obj(np.load('/workspace/best_n2000_ultrahigh.npy')):
    np.save('/workspace/best_tiled.npy', best_a)
    print("NEW RECORD!")