Scipy全局优化中Differential Evolution参数传递报错求助

Let's break down what's causing your error and how to fix it step by step:

1. Root Cause of the Error #

Your callback function is calling objective(x) without passing the args parameter (the load value). When the callback runs, objective receives an empty *args, so np.append(x, args) just returns the original 2-element x array. Your ML model expects 3 input parameters, which leads to a shape mismatch when calling MLmodel.predict(x)—hence the broadcast together error.

Additionally, depending on your ML model's input requirements, the 1D array from np.append might need to be reshaped to a 2D array (since most scikit-learn-style models expect inputs in (n_samples, n_features) format).

2. Corrected Code #

Here's the fixed version with explanations:

import numpy as np
from scipy import optimize

def objective(x, *args):
    # Combine optimized parameters with fixed argument, reshape to match ML model input
    full_params = np.append(x, args).reshape(1, -1)
    res = MLmodel.predict(full_params)
    return res[0]  # Extract scalar (predict returns array even for single sample)

fun_history = []
x_values = []
load = (50,)  # Fixed third parameter (keep as tuple for Scipy's args)

def callback(x, convergence):
    # Pass the fixed args to objective when tracking history
    fobj = objective(x, *load)
    x_values.append(x.copy())  # Copy to avoid overwriting by DE's internal array reuse
    fun_history.append(fobj)

bounds = [(5.5,8.8),(29,40)]
res = optimize.differential_evolution(
    objective,
    bounds,
    args=load,
    disp=True,
    callback=callback
)

3. Key Modifications Explained #

• Callback Argument Fix: We now pass *load to objective inside the callback, ensuring the fixed third parameter is included when calculating the objective value for history tracking.
• Input Reshaping: Added .reshape(1, -1) to convert the combined 1D parameter array into a 2D array, which aligns with the input format most ML models expect (single sample, 3 features).
• Scalar Extraction: Used return res[0] because predict typically returns an array (even for a single sample), and Differential Evolution requires a scalar objective value to optimize.
• Copying x Values: Added x.copy() when appending to x_values—Differential Evolution reuses and modifies the same array internally, so without copying, your history will show the final x value repeated instead of each iteration's unique values.

4. Quick Additional Checks #

If you still run into shape issues:

• Verify your MLmodel's expected input shape (use MLmodel.n_features_in_ if it's a scikit-learn model).
• Double-check that load is a tuple (even for a single value, like (50,) instead of 50—Scipy's args parameter requires a tuple input).

内容的提问来源于stack exchange，提问作者chink