ccevolve/tx_workspace/ac1_alphaevolve/solution_v17.py

"""
Optimize step function heights using JAX.
K parameters (step heights) instead of N parameters.
Lower-dimensional search = better global exploration.
"""
import sys
import jax
import jax.numpy as jnp
import numpy as np
from scipy.optimize import minimize as scipy_minimize, differential_evolution
import optax


def compute_c1_numpy(f_values, n_points):
    dx = 0.5 / n_points
    f_nn = np.maximum(f_values, 0.0)
    autoconv = np.convolve(f_nn, f_nn, mode='full') * dx
    integral_sq = (np.sum(f_nn) * dx) ** 2
    if integral_sq < 1e-12:
        return 1e10
    return np.max(autoconv) / integral_sq


def make_step_fns(K, N):
    """Create functions for K-step function optimization at N resolution."""
    dx = 0.5 / N
    # Precompute indices for each step
    step_starts = jnp.array([int(i * N / K) for i in range(K)])
    step_ends = jnp.array([int((i + 1) * N / K) for i in range(K)])

    # Create a mapping matrix: f[j] = heights[i] if j belongs to step i
    mapping = jnp.zeros((K, N))
    for i in range(K):
        s = int(i * N / K)
        e = int((i + 1) * N / K)
        mapping = mapping.at[i, s:e].set(1.0)

    @jax.jit
    def heights_to_f(log_heights):
        heights = jnp.exp(jnp.clip(log_heights, -4, 4))
        f = jnp.dot(heights, mapping)
        return f

    @jax.jit
    def obj_smooth(log_heights, temp):
        f = heights_to_f(log_heights)
        padded = jnp.zeros(2 * N)
        padded = padded.at[:N].set(f)
        fft_f = jnp.fft.rfft(padded)
        conv = jnp.fft.irfft(fft_f * fft_f, n=2 * N) * dx
        integral_sq = (jnp.sum(f) * dx) ** 2
        smooth_max = jax.nn.logsumexp(temp * conv) / temp
        return smooth_max / integral_sq

    @jax.jit
    def obj_hard(log_heights):
        f = heights_to_f(log_heights)
        padded = jnp.zeros(2 * N)
        padded = padded.at[:N].set(f)
        fft_f = jnp.fft.rfft(padded)
        conv = jnp.fft.irfft(fft_f * fft_f, n=2 * N) * dx
        integral_sq = (jnp.sum(f) * dx) ** 2
        return jnp.max(conv) / integral_sq

    grad_smooth = jax.jit(jax.grad(obj_smooth))
    grad_hard = jax.jit(jax.grad(obj_hard))

    return heights_to_f, obj_smooth, obj_hard, grad_smooth, grad_hard


def optimize_step(K, N, seed=42, adam_steps=100000):
    heights_to_f, obj_smooth, obj_hard, grad_smooth, grad_hard = make_step_fns(K, N)

    np.random.seed(seed)
    # Start near constant function
    log_heights = jnp.array(np.random.randn(K) * 0.1)

    # Adam optimization
    lr_schedule = optax.warmup_cosine_decay_schedule(
        init_value=0.0, peak_value=0.01, warmup_steps=1000,
        decay_steps=adam_steps - 1000, end_value=1e-6,
    )
    optimizer = optax.adam(learning_rate=lr_schedule)
    opt_state = optimizer.init(log_heights)

    best_c1 = float('inf')
    best_params = log_heights

    for step in range(adam_steps):
        temp = 300.0
        loss, grads = jax.value_and_grad(obj_smooth)(log_heights, temp)
        updates, opt_state = optimizer.update(grads, opt_state, log_heights)
        log_heights = optax.apply_updates(log_heights, updates)

        if step % 25000 == 0 or step == adam_steps - 1:
            hc = float(obj_hard(log_heights))
            if hc < best_c1:
                best_c1 = hc
                best_params = log_heights

    # L-BFGS polish
    params_np = np.array(best_params, dtype=np.float64)

    for temp in [1000.0, 10000.0]:
        def scipy_obj(p):
            p_jax = jnp.array(p)
            val = float(obj_smooth(p_jax, temp))
            g = np.array(grad_smooth(p_jax, temp), dtype=np.float64)
            return val, g
        result = scipy_minimize(
            scipy_obj, params_np, method='L-BFGS-B', jac=True,
            options={'maxiter': 10000, 'ftol': 1e-15, 'gtol': 1e-14},
        )
        params_np = result.x

    # Hard max L-BFGS
    def scipy_obj_hard(p):
        p_jax = jnp.array(p)
        val = float(obj_hard(p_jax))
        g = np.array(grad_hard(p_jax), dtype=np.float64)
        return val, g
    result = scipy_minimize(
        scipy_obj_hard, params_np, method='L-BFGS-B', jac=True,
        options={'maxiter': 20000, 'ftol': 1e-16, 'gtol': 1e-15},
    )
    params_np = result.x

    f = np.array(heights_to_f(jnp.array(params_np)))
    c1 = compute_c1_numpy(f, N)
    return f, c1, params_np


def run():
    N = 4000
    best_c1 = float('inf')
    best_f = None

    # Try many different numbers of steps
    for K in [20, 30, 40, 50, 60, 80, 100, 150, 200, 300, 500, 800, 1000, 2000]:
        best_k_c1 = float('inf')
        for seed in range(5):
            f, c1, params = optimize_step(K, N, seed=seed, adam_steps=50000)
            if c1 < best_k_c1:
                best_k_c1 = c1

            if c1 < best_c1:
                best_c1 = c1
                best_f = f
                sys.stdout.write(f"K={K:4d} seed={seed}: C1={c1:.10f} ***\n")
            else:
                if seed == 0 and c1 < best_k_c1 + 0.001:
                    sys.stdout.write(f"K={K:4d} seed={seed}: C1={c1:.10f}\n")
            sys.stdout.flush()

    # Also try full resolution (K=N) with best seeds
    K_full = N
    sys.stdout.write(f"\nFull resolution K={K_full}...\n")
    sys.stdout.flush()
    for seed in [15, 8, 2]:  # best seeds from v16
        f, c1, _ = optimize_step(K_full, N, seed=seed, adam_steps=80000)
        sys.stdout.write(f"K={K_full} seed={seed}: C1={c1:.10f}\n")
        sys.stdout.flush()
        if c1 < best_c1:
            best_c1 = c1
            best_f = f
            sys.stdout.write(f"*** NEW GLOBAL BEST\n")
            sys.stdout.flush()

    sys.stdout.write(f"\nFinal C1: {best_c1:.10f}\n")
    sys.stdout.flush()
    return best_f, best_c1, best_c1, N