utils.py 11.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
# -*- coding: utf-8 -*-
"""
Created on Sun Aug 18 14:30:09 2013

@author: steve
"""

from numpy import absolute, ones

10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
# These need to be fixed so that we use classes derived from Error.
mdperr = {
"mat_nonneg" :
    "Probabilities must be non-negative.",
"mat_square" :
    "The matrix must be square.",
"mat_stoch" :
    "Rows of the matrix must sum to one (1).",
"mask_numpy" :
    "mask must be a numpy array or matrix; i.e. type(mask) is "
    "ndarray or type(mask) is matrix.", 
"mask_SbyS" : 
    "The mask must have shape SxS; i.e. mask.shape = (S, S).",
"obj_shape" :
    "Object arrays for transition probabilities and rewards "
    "must have only 1 dimension: the number of actions A. Each element of "
    "the object array contains an SxS ndarray or matrix.",
"obj_square" :
    "Each element of an object array for transition "
    "probabilities and rewards must contain an SxS ndarray or matrix; i.e. "
    "P[a].shape = (S, S) or R[a].shape = (S, S).",
"P_type" :
    "The transition probabilities must be in a numpy array; "
    "i.e. type(P) is ndarray.",
"P_shape" :
    "The transition probability array must have the shape "
    "(A, S, S)  with S : number of states greater than 0 and A : number of "
    "actions greater than 0. i.e. R.shape = (A, S, S)",
"PR_incompat" :
    "Incompatibility between P and R dimensions.",
"prob_in01" :
    "Probability p must be in [0; 1].",
"R_type" :
    "The rewards must be in a numpy array; i.e. type(R) is "
    "ndarray, or numpy matrix; i.e. type(R) is matrix.",
"R_shape" :
    "The reward matrix R must be an array of shape (A, S, S) or "
    "(S, A) with S : number of states greater than 0 and A : number of "
    "actions greater than 0. i.e. R.shape = (S, A) or (A, S, S).",
"R_gt_0" :
    "The rewards must be greater than 0.",
"S_gt_1" :
    "Number of states S must be greater than 1.",
"SA_gt_1" : 
    "The number of states S and the number of actions A must be "
    "greater than 1."
}

58
def check(P, R):
59
    """Check if P and R define a valid Markov Decision Process (MDP).
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
    
    Let S = number of states, A = number of actions.
    
    Parameters
    ---------
    P : array
        The transition matrices. It can be a three dimensional array with
        a shape of (A, S, S). It can also be a one dimensional arraye with
        a shape of (A, ), where each element contains a matrix of shape (S, S)
        which can possibly be sparse.
    R : array
        The reward matrix. It can be a three dimensional array with a
        shape of (S, A, A). It can also be a one dimensional array with a
        shape of (A, ), where each element contains matrix with a shape of
        (S, S) which can possibly be sparse. It can also be an array with
        a shape of (S, A) which can possibly be sparse.  
    
    Notes
    -----
    Raises an error if P and R do not define a MDP.
80
81
82
83
84
85
86
87
88
    
    Examples
    --------
    >>> import mdptoolbox, mdptoolbox.example
    >>> P_valid, R_valid = mdptoolbox.example.rand(100, 5)
    >>> mdptoolbox.utils.check(P_valid, R_valid) # Nothing should happen
    >>> 
    >>> import numpy as np
    >>> P_invalid = np.random.rand(5, 100, 100)
89
    >>> mdptoolbox.utils.check(P_invalid, R_valid) # Raises an exception
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290

    """
    # Checking P
    try:
        if P.ndim == 3:
            aP, sP0, sP1 = P.shape
        elif P.ndim == 1:
            # A hack so that we can go into the next try-except statement and
            # continue checking from there
            raise AttributeError
        else:
            raise ValueError(mdperr["P_shape"])
    except AttributeError:
        try:
            aP = len(P)
            sP0, sP1 = P[0].shape
            for aa in xrange(1, aP):
                sP0aa, sP1aa = P[aa].shape
                if (sP0aa != sP0) or (sP1aa != sP1):
                    raise ValueError(mdperr["obj_square"])
        except AttributeError:
            raise TypeError(mdperr["P_shape"])
    except:
        raise
    # Checking R
    try:
        if R.ndim == 2:
            sR0, aR = R.shape
            sR1 = sR0
        elif R.ndim == 3:
            aR, sR0, sR1 = R.shape
        elif R.ndim == 1:
            # A hack so that we can go into the next try-except statement
            raise AttributeError
        else:
            raise ValueError(mdperr["R_shape"])
    except AttributeError:
        try:
            aR = len(R)
            sR0, sR1 = R[0].shape
            for aa in range(1, aR):
                sR0aa, sR1aa = R[aa].shape
                if ((sR0aa != sR0) or (sR1aa != sR1)):
                    raise ValueError(mdperr["obj_square"])
        except AttributeError:
            raise ValueError(mdperr["R_shape"])
    except:
        raise
    # Checking dimensions
    if (sP0 < 1) or (aP < 1) or (sP0 != sP1):
        raise ValueError(mdperr["P_shape"])   
    if (sR0 < 1) or (aR < 1) or (sR0 != sR1):
        raise ValueError(mdperr["R_shape"])
    if (sP0 != sR0) or (aP != aR):
        raise ValueError(mdperr["PR_incompat"])
    # Check that the P's are square and stochastic
    for aa in xrange(aP):
        checkSquareStochastic(P[aa])
        #checkSquareStochastic(P[aa, :, :])
    # We are at the end of the checks, so if no exceptions have been raised
    # then that means there are (hopefullly) no errors and we return None
    return None
    
    # These are the old code comments, which need to be converted to
    # information in the docstring:
    #
    # tranitions must be a numpy array either an AxSxS ndarray (with any 
    # dtype other than "object"); or, a 1xA ndarray with a "object" dtype, 
    # and each element containing an SxS array. An AxSxS array will be
    # be converted to an object array. A numpy object array is similar to a
    # MATLAB cell array.
    #
    # NumPy has an array type of 'object', which is roughly equivalent to
    # the MATLAB cell array. These are most useful for storing sparse
    # matrices as these can only have two dimensions whereas we want to be
    # able to store a transition matrix for each action. If the dytpe of
    # the transition probability array is object then we store this as
    # P_is_object = True.
    # If it is an object array, then it should only have one dimension
    # otherwise fail with a message expalining why.
    # If it is a normal array then the number of dimensions must be exactly
    # three, otherwise fail with a message explaining why.
    #
    # As above but for the reward array. A difference is that the reward
    # array can have either two or 3 dimensions.
    #
    # We want to make sure that the transition probability array and the 
    # reward array are in agreement. This means that both should show that
    # there are the same number of actions and the same number of states.
    # Furthermore the probability of transition matrices must be SxS in
    # shape, so we check for that also.
    #
        # If the user has put their transition matrices into a numpy array
        # with dtype of 'object', then it is possible that they have made a
        # mistake and not all of the matrices are of the same shape. So,
        # here we record the number of actions and states that the first
        # matrix in element zero of the object array says it has. After
        # that we check that every other matrix also reports the same
        # number of actions and states, otherwise fail with an error.
        # aP: the number of actions in the transition array. This
        # corresponds to the number of elements in the object array.
        #
        # sP0: the number of states as reported by the number of rows of
        # the transition matrix
        # sP1: the number of states as reported by the number of columns of
        # the transition matrix
        #
        # Now we check to see that every element of the object array holds
        # a matrix of the same shape, otherwise fail.
        #
            # sp0aa and sp1aa represents the number of states in each
            # subsequent element of the object array. If it doesn't match
            # what was found in the first element, then we need to fail
            # telling the user what needs to be fixed.
            #
        # if we are using a normal array for this, then the first
        # dimension should be the number of actions, and the second and 
        # third should be the number of states
        #
    # the first dimension of the transition matrix must report the same
    # number of states as the second dimension. If not then we are not
    # dealing with a square matrix and it is not a valid transition
    # probability. Also, if the number of actions is less than one, or the
    # number of states is less than one, then it also is not a valid
    # transition probability.
    #
    # now we check that each transition matrix is square-stochastic. For
    # object arrays this is the matrix held in each element, but for
    # normal arrays this is a matrix formed by taking a slice of the array
    #
        # if the rewarad array has an object dtype, then we check that
        # each element contains a matrix of the same shape as we did 
        # above with the transition array.
        #
        # This indicates that the reward matrices are constructed per 
        # transition, so that the first dimension is the actions and
        # the second two dimensions are the states.
        #
        # then the reward matrix is per state, so the first dimension is 
        # the states and the second dimension is the actions.
        #
        # this is added just so that the next check doesn't error out
        # saying that sR1 doesn't exist
        #
    # the number of actions must be more than zero, the number of states
    # must also be more than 0, and the states must agree
    #
    # now we check to see that what the transition array is reporting and
    # what the reward arrar is reporting agree as to the number of actions
    # and states. If not then fail explaining the situation

def checkSquareStochastic(Z):
    """Check if Z is a square stochastic matrix.
    
    Let S = number of states.
    
    Parameters
    ----------
    Z : matrix
        This should be a two dimensional array with a shape of (S, S). It can
        possibly be sparse.
    
    Notes 
    ----------
    Returns None if no error has been detected, else it raises an error.
    
    """
    # try to get the shape of the matrix
    try:
        s1, s2 = Z.shape
    except AttributeError:
        raise TypeError("Matrix should be a numpy type.")
    except ValueError:
        raise ValueError(mdperr["mat_square"])
    # check that the matrix is square, and that each row sums to one
    if s1 != s2:
        raise ValueError(mdperr["mat_square"])
    elif (absolute(Z.sum(axis=1) - ones(s2))).max() > 10e-12:
        raise ValueError(mdperr["mat_stoch"])
    # make sure that there are no values less than zero
    try:
        if (Z < 0).any():
            raise ValueError(mdperr["mat_nonneg"])
    except AttributeError:
        try:
            if (Z.data < 0).any():
                raise ValueError(mdperr["mat_nonneg"])
        except AttributeError:
            raise TypeError("Matrix should be a numpy type.")
    except:
        raise
    
    return(None)

def getSpan(W):
    """Return the span of W
    
    sp(W) = max W(s) - min W(s)
    
    """
    return (W.max() - W.min())
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308

class Error(Exception):
    """Base class for exceptions in this module."""
    
    def __init__(self):
        Exception.__init__(self)
        self.message = "PyMDPToolbox: "
    
    def __str__(self):
        return repr(self.message)

class InvalidMDPError(Error):
    """Class for invalid definitions of a MDP."""
    
    def __init__(self, msg):
        Error.__init__(self)
        self.message += msg
        self.args = tuple(msg)