Polytope error for 1D - Githubissues

shaesaert commented 7 years ago

In one dimensional case studies, the abstraction algorithm of TuLiP triggers a compatibility error. This can be fixed easily: Polytope file:

if n==1:
    x = [sol['x']]
else:
    x = sol['x']

lines 1285 to 1289 lines 1297 to 1300

johnyf commented 7 years ago

Thanks for reporting this error. This error does not arise when scipy is present and cvxopt.glpk absent. It arises with cvxopt.glpk present. Minimal example using e030e125108ff6cd4ffa1ffdb199ffc9864a4b1f:

In [1]: import cvxopt.glpk

In [2]: import polytope as pt

In [3]: p = pt.Polytope.from_box([[0, 1]])

In [4]: p.bounding_box
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-4-03e06b93ceaf> in <module>()
----> 1 p.bounding_box

/path/polytope/polytope/polytope.py in bounding_box(self)
    406         """
    407         if self.bbox is None:
--> 408             self.bbox = bounding_box(self)
    409         return self.bbox
    410 

/path/path/polytope/polytope/polytope.py in bounding_box(polyreg)
   1288         if sol['status'] == 0:
   1289             x = sol['x']
-> 1290             l[i] = x[i]
   1291     # upper corner
   1292     for i in xrange(0, n):

IndexError: too many indices for array

The relevant lines are polytope.polytope#L1290 and polytope.polytope#L1299 (corresponding to lower and upper corners of the bounding box, respectively).

This bug was introduced in 79d85ec95a32819534011756bb0200c824ae2148. It does not exist in 25dbe65eabb89ed6f9a4450a822489558438708f and appears in the merge commit 907934a2a698d7e77c1cd5334bb9c9e5cb18ebf0. The bug is caused by a zero-dimensional array returned by the function lpsolve.

The error is introduced by the call to numpy.squeeze on line 2282. Squeezing yields a zero dimensional array when fed with an array of shape (1,):

import numpy as np

a = np.array(1)
>>> a.shape
()
>>> a.ndim
()

b = np.array([1])
>>> b.shape
(1,)

c = np.squeeze(b)
>>> c.shape
()

In other words, there is still an asymmetry between the GLPK branch and the SciPy branch of the selection statement within lpsolve. So the aim of 79d85ec95a32819534011756bb0200c824ae2148 has not been accomplished yet.

johnyf commented 6 years ago

tulip-control / polytope

Polytope error for 1D #47