ccevolve/tx_workspace/ac1_alphaevolve/solution_v6.py

import jax
import jax.numpy as jnp
import numpy as np
from scipy.optimize import minimize as scipy_minimize
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_fns(N):
    dx = 0.5 / N

    @jax.jit
    def get_f(params):
        return jnp.exp(jnp.clip(params, -8, 4))

    @jax.jit
    def objective_smooth(params, temp):
        f = get_f(params)
        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 objective_hard(params):
        f = get_f(params)
        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(objective_smooth))

    return get_f, objective_smooth, objective_hard, grad_smooth


def optimize(N, adam_steps, lr_peak, seed=42, init_params=None, verbose=True):
    dx = 0.5 / N
    get_f, objective_smooth, objective_hard, grad_smooth = make_fns(N)

    if init_params is not None:
        # Upsample
        old_f = np.exp(np.clip(init_params, -8, 4))
        new_f = np.interp(np.linspace(0, 1, N), np.linspace(0, 1, len(init_params)), old_f)
        params = jnp.array(np.log(np.maximum(new_f, 1e-6)))
    else:
        np.random.seed(seed)
        init_f = np.ones(N) * 0.5 + 0.02 * np.random.randn(N)
        params = jnp.array(np.log(np.maximum(init_f, 1e-6)))

    # Adam optimization with fixed moderate temperature
    lr_schedule = optax.warmup_cosine_decay_schedule(
        init_value=0.0, peak_value=lr_peak, warmup_steps=2000,
        decay_steps=adam_steps - 2000, end_value=lr_peak * 1e-5,
    )
    optimizer = optax.adam(learning_rate=lr_schedule)
    opt_state = optimizer.init(params)

    best_c1 = float('inf')
    best_params = params
    temp = 200.0

    for step in range(adam_steps):
        loss, grads = jax.value_and_grad(objective_smooth)(params, temp)
        # Clip gradients
        grads = jnp.clip(grads, -1.0, 1.0)
        updates, opt_state = optimizer.update(grads, opt_state, params)
        params = optax.apply_updates(params, updates)

        if step % 10000 == 0 or step == adam_steps - 1:
            hard_c1 = float(objective_hard(params))
            if verbose:
                print(f"  Step {step:6d} | C1={hard_c1:.8f}")
            if hard_c1 < best_c1:
                best_c1 = hard_c1
                best_params = params

    # L-BFGS polishing
    if verbose:
        print(f"  L-BFGS polishing from C1={best_c1:.8f}")

    params_np = np.array(best_params, dtype=np.float64)
    for temp_lbfgs in [500.0, 2000.0, 5000.0, 20000.0]:
        def scipy_obj(p):
            p_jax = jnp.array(p)
            val = float(objective_smooth(p_jax, temp_lbfgs))
            g = np.array(grad_smooth(p_jax, temp_lbfgs), dtype=np.float64)
            return val, g

        result = scipy_minimize(
            scipy_obj, params_np, method='L-BFGS-B', jac=True,
            options={'maxiter': 5000, 'ftol': 1e-15, 'gtol': 1e-12},
        )
        params_np = result.x
        f_opt = np.exp(np.clip(params_np, -8, 4))
        c1 = compute_c1_numpy(f_opt, N)
        if verbose:
            print(f"    temp={temp_lbfgs:.0f}: C1={c1:.10f}")
        if c1 < best_c1:
            best_c1 = c1
            best_params = jnp.array(params_np)

    return np.array(best_params), best_c1


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

    # Stage 1: coarse
    print("=== Stage 1: N=1500 ===")
    params, c1 = optimize(1500, 80000, 0.008, seed=42)
    print(f"  Stage 1 C1: {c1:.10f}")

    # Stage 2: medium
    print("\n=== Stage 2: N=3000 ===")
    params, c1 = optimize(3000, 80000, 0.004, init_params=params)
    print(f"  Stage 2 C1: {c1:.10f}")

    # Stage 3: fine
    print("\n=== Stage 3: N=5000 ===")
    params, c1 = optimize(5000, 60000, 0.002, init_params=params)
    print(f"  Stage 3 C1: {c1:.10f}")

    N = 5000
    f_final = np.exp(np.clip(params, -8, 4))
    c1_final = compute_c1_numpy(f_final, N)

    if c1_final < best_c1:
        best_c1 = c1_final
        best_f = f_final
        best_n = N

    # Also try direct N=4000 from scratch with different seed
    print("\n=== Direct: N=4000 seed=0 ===")
    params2, c1_2 = optimize(4000, 100000, 0.005, seed=0)
    print(f"  Direct C1: {c1_2:.10f}")
    if c1_2 < best_c1:
        N2 = 4000
        f2 = np.exp(np.clip(params2, -8, 4))
        best_c1 = c1_2
        best_f = f2
        best_n = N2

    print(f"\nFinal best C1: {best_c1:.10f}")
    return best_f, best_c1, best_c1, best_n