scipy.optimize.milp未遵守time_limit参数，超时后长时间无法返回的问题求助

数据老陈

2026-4-7

大家好，我最近在用scipy的milp函数求解混合整数线性规划问题，遇到了一个特别头疼的问题：明明已经设置了time_limit参数（比如我设的是540秒），而且求解报告已经打印出来了（不管是找到最优解还是触发时间限制），但函数就是不立刻返回，要等很久（甚至1500秒）才会继续执行后面的代码，比如我在函数调用后写的打印耗时的语句。哪怕是处理小数据集，虽然返回速度会快一些，但还是会出现这种“求解报告出来后挂起”的情况。

先贴一下我看到的求解报告内容：

Status Optimal
Primal bound -39
Dual bound -39
Gap 0% (tolerance: 0.01%)
P-D integral 0
Solution status feasible
-39 (objective)
0 (bound viol.)
0 (int. viol.)
0 (row viol.)
Timing
12.29 (total)
0.00 (presolve)
0.00 (solve)
0.00 (postsolve)
Max sub-MIP depth 12
Nodes 1
Repair LPs 0 (0 feasible; 0 iterations)
LP iterations 29704 (total)
0 (strong br.)
2544 (separation)
18467 (heuristics)

我已经尝试过简化输入数据来降低求解难度，确实小数据集的返回速度会快一些，但哪怕是小数据集，还是会出现求解完成后挂起的情况。我当前使用的scipy版本是1.15.2。

下面是我初始化和调用MILP的核心代码：

def solve_and_extract(df_chunk, time_limit):
    n = len(df_chunk)
    if n < 10:
        return [], df_chunk.copy(), False
    m = n // 10
    
    caps = df_chunk['Col1'].astype(float).to_numpy()
    fr = df_chunk['Col2'].astype(float).to_numpy()
    
    c = np.concatenate([np.zeros(n*m), -np.ones(m), np.zeros(n)])
    integrality = np.ones(n*m + m + n, dtype=int)
    bounds = Bounds(0, 1)
    
    A_link1 = np.zeros((n, n*m + m + n)); b_link1 = np.zeros(n)
    A_link2 = np.zeros((n, n*m + m + n)); b_link2 = np.zeros(n)
    for i in range(n):
        A_link1[i, i + np.arange(m)*n] = 1
        A_link1[i, n*m + m + i] = -1
        A_link2[i, i + np.arange(m)*n] = -1
        A_link2[i, n*m + m + i] = 1
    
    A2a = np.zeros((m, n*m + m + n)); b2a = np.zeros(m)
    A2b = np.zeros((m, n*m + m + n)); b2b = np.zeros(m)
    for j in range(m):
        A2a[j, j*n:(j+1)*n] = 1
        A2a[j, n*m + j] = -10
        A2b[j, j*n:(j+1)*n] = -1
        A2b[j, n*m + j] = 10
    
    A3 = np.zeros((m, n*m + m + n)); b3 = np.zeros(m)
    A4 = np.zeros((m, n*m + m + n)); b4 = np.zeros(m)
    A5 = np.zeros((m, n*m + m + n)); b5 = np.zeros(m)
    A6 = np.zeros((m, n*m + m + n)); b6 = np.zeros(m)
    for j in range(m):
        A3[j, j*n:(j+1)*n] = caps
        A3[j, n*m + j] = -maxc*10
        A4[j, j*n:(j+1)*n] = -caps
        A4[j, n*m + j] = minc*10
        A5[j, j*n:(j+1)*n] = fr
        A5[j, n*m + j] = -max_fr*10
        A6[j, j*n:(j+1)*n] = -fr
        A6[j, n*m + j] = min_fr*10
    
    A_sym = np.zeros((m-1, n*m + m + n)); b_sym = np.zeros(m-1)
    for j in range(m-1):
        A_sym[j, n*m + j] = 1
        A_sym[j, n*m + j+1] = -1
    
    A7 = np.zeros((1, n*m + m + n))
    A7[0, n*m + m:] = +1
    A7[0, n*m:n*m+m] = -10
    
    constraints = [
        LinearConstraint(A2a, -np.inf, b2a),
        LinearConstraint(A2b, -np.inf, b2b),
        LinearConstraint(A3, -np.inf, b3),
        LinearConstraint(A4, -np.inf, b4),
        LinearConstraint(A5, -np.inf, b5),
        LinearConstraint(A6, -np.inf, b6),
        LinearConstraint(A7, -np.inf, 0),
        LinearConstraint(A_link1, -np.inf, b_link1),
        LinearConstraint(A_link2, -np.inf, b_link2),
        LinearConstraint(A_sym, b_sym, np.inf)
    ]
    
    res = milp(
        c=c,
        integrality=integrality,
        bounds=bounds,
        constraints=constraints,
        options={"disp": True, "time_limit": time_limit}
    )
    # 后续处理逻辑...

请问各位，有没有办法优化这种情况？为什么milp函数在求解报告输出后还会挂起这么久？是我的代码结构有问题，还是scipy的MILP求解器本身存在潜在问题？有没有可以调整的参数或者技巧，能让函数在求解完成后立刻返回？

内容来源于stack exchange