Question

I have a function in python which takes a vector and returns a real number. I am using the scipy.optimize fmin and fmin_bfgs functions to find the argument which gives the function its approx minimum value. However, when I use fmin I get an alright answer (quite slowly) but when I switch to fmin_bfgs, I get an error saying “Matrices are not aligned”. Here’s my function:

    def norm(b_):
        b_ = b_.reshape(int(M),1)                   #M already given elsewhere
        Yb = np.dot(Y,b_)                           #Y already given elsewhere

        B = np.zeros((int(M),int(M))) 
        for j in xrange(int(M)):
            B[j][j] = -t[j+1]*np.exp(-t[j+1]*Yb[j]) #The t[j] are already known
        P = np.zeros((int(M),1))
        for j in xrange(int(M)):
            P[j][0] = np.exp(-t[j+1]*Yb[j])

        diff = np.zeros((int(M),1))                 #Functions d(i,b) are known
        for i in xrange(1,int(M)-1):
            diff[i][0] = d(i+1,b_) - d(i,b_)
        diff[0][0] = d(1,b_)
        diff[int(M)-1][0] = -d(int(M)-1,b_)

        term1_ = (1.0/N)*(np.dot((V - np.dot(c,P)).transpose(),W))
        term2_ = np.dot(W,V - np.dot(c,P))          #V,c,P,W already known
        term1_ = np.dot(term1_,term2_)
        term2_ = lambd*np.dot(Yb.transpose(),diff)

        return term1_ + term2_

Here’s how I call fmin_bfgs:

    fmin_bfgs(norm, b_guess,fprime=None,
              args=(),gtol=0.0001,norm=0.00000000001,
              epsilon=1.4901161193847656e-08,maxiter=None,
              full_output=0, disp=1, retall=0, callback=None)

When I call fmin it works fine, just too slowly to be useful (I need to optimise several times). But when I try fmin_bfgs I get this error:

Traceback (most recent call last):
File “C:\Program Files\Wing IDE 101 4.0\src\debug\tserver_sandbox.py”, line 287, in module
File “C:\Python27\Lib\site-packages\scipy\optimize\optimize.py”, line 491, in fmin_bfgs old_fval,old_old_fval)
File “C:\Python27\Lib\site-packages\scipy\optimize\linesearch.py”, line 239, in line_search_wolfe2 derphi0, c1, c2, amax)
File “C:\Python27\Lib\site-packages\scipy\optimize\linesearch.py”, line 339, in scalar_search_wolfe2 phi0, derphi0, c1, c2)
File “C:\Python27\Lib\site-packages\scipy\optimize\linesearch.py”, line 471, in _zoom derphi_aj = derphi(a_j)
File “C:\Python27\Lib\site-packages\scipy\optimize\linesearch.py”, line 233, in derphi return np.dot(gval[0], pk)
ValueError: matrices are not aligned

Any ideas why this might happen? All the matrices I have supplied the function are aligned correctly (and the function works since fmin worked). Help much appreciated!

Answer 1

It seems that one of the programs just ended up dealing with numbers that were too large for it to handle. Shame it couldn’t tell me it was doing that properly. I worked around it though, so no more problem. Sorry if this wasted your time.