函数scipy.optimize.minimize的选项

Question

I am trying to minimize a very long function (it is the sum of 500000 parts of sub-functions) in order to fit some parameters to a probabilistic model. I use the scipy.optimize.minimize function. I tried both Powell and Nelder-Mead algorithms, and Powell looks really faster in my settings. But still, I really don't understand how to force the process to give me some results after a given time, even if they are not "optimal".

I fill the options maxiter , maxfev , xtol and ftol , but I don't really understand these options, since I tried to put a print in my function and I noticed that the algorithm evaluate it more than maxfev times, but when It reaches the maxiter point, It sends an error "max number of iterations reached".

Can anyone explain me how they work with respect to the two algorithms I am using? The doc is very unclear.

My code:

def log_likelihood(r, alpha, a, b, customers):
    if r <= 0 or alpha <= 0 or a <= 0 or b <= 0:
        return -np.inf
    c = sum([log_likelihood_individual(r, alpha, a, b, x, tx, t) for x, tx, t in customers])
    print -c
    return c

negative_ll = lambda params: -log_likelihood(*params,customers=customers)
params0 = (1, 1, 1, 1)
res = minimize(negative_ll, params0, method='Powell', callback=print_callback, options={'disp': True, 'ftol':0.05, 'maxiter':3, 'maxfev":15})

Thank you.

Answer 1

You should probably ask this on the scipy mailing list, or even the scipy developer mailing list, but looking at the source code for the Nelder-Mead algorithm, I notice the actual check on maxiter and maxfev are in the outer while loop. The function is called several times inside that while loop, so the actual number of function evaluations can easily surpass maxfev . Something similar happens inside the main loop for Powell's method. For that method, it appears the function is evaluated N times before the number of evaluations is tested ( N the number of parameters).

I guess this is done because otherwise there would be too many if statements inside the core loops that check against maxfev , and it was deemed faster/clearer to have the condition outside the inner loops.