def _checkRewardsListLike(reward, n_actions, n_states):
"""Check that a list-like reward input is valid.
"""
try:
lenR = len(reward)
if lenR == n_actions:
dim1, dim2, dim3 = _checkDimensionsListLike(reward)
elif lenR == n_states:
dim1 = n_actions
dim2 = dim3 = lenR
else:
raise _error.InvalidError(_MDPERR["R_shape"])
except AttributeError:
raise _error.InvalidError(_MDPERR["R_shape"])
return dim1, dim2, dim3
def _checkDimensionsListLike(arrays):
"""Check that each array in a list of arrays has the same size.
"""
dim1 = len(arrays)
dim2, dim3 = arrays[0].shape
for aa in range(1, dim1):
dim2_aa, dim3_aa = arrays[aa].shape
if (dim2_aa != dim2) or (dim3_aa != dim3):
raise _error.InvalidError(_MDPERR["obj_square"])
return dim1, dim2, dim3
def check(P, R):
"""Check if ``P`` and ``R`` define a valid Markov Decision Process (MDP).
Let ``S`` = number of states, ``A`` = number of actions.
Arguments
---------
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.
Examples
--------
>>> import mdptoolbox, mdptoolbox.example
>>> P_valid, R_valid = mdptoolbox.example.rand(100, 5)
>>> mdptoolbox.util.check(P_valid, R_valid) # Nothing should happen
>>>
>>> import numpy as np
>>> P_invalid = np.random.rand(5, 100, 100)
>>> mdptoolbox.util.check(P_invalid, R_valid) # Raises an exception
Traceback (most recent call last):
...
StochasticError:...
"""
# Checking P
try:
if P.ndim == 3:
aP, sP0, sP1 = P.shape
elif P.ndim == 1:
aP, sP0, sP1 = _checkDimensionsListLike(P)
else:
raise _error.InvalidError(_MDPERR["P_shape"])
except AttributeError:
try:
aP, sP0, sP1 = _checkDimensionsListLike(P)
except AttributeError:
raise _error.InvalidError(_MDPERR["P_shape"])
msg = ""
if aP <= 0:
msg = "The number of actions in P must be greater than 0."
elif sP0 <= 0:
msg = "The number of states in P must be greater than 0."
if msg:
raise _error.InvalidError(msg)
# Checking R
try:
ndimR = R.ndim
if ndimR == 1:
aR, sR0, sR1 = _checkRewardsListLike(R, aP, sP0)
elif ndimR == 2:
sR0, aR = R.shape
sR1 = sR0
elif ndimR == 3:
aR, sR0, sR1 = R.shape
else:
raise _error.InvalidError(_MDPERR["R_shape"])
except AttributeError:
aR, sR0, sR1 = _checkRewardsListLike(R, aP, sP0)
msg = ""
if sR0 <= 0:
msg = "The number of states in R must be greater than 0."
elif aR <= 0:
msg = "The number of actions in R must be greater than 0."
elif sR0 != sR1:
msg = "The matrix R must be square with respect to states."
elif sP0 != sR0:
msg = "The number of states must agree in P and R."
elif aP != aR:
msg = "The number of actions must agree in P and R."
if msg:
raise _error.InvalidError(msg)
# Check that the P's are square, stochastic and non-negative
for aa in range(aP):
checkSquareStochastic(P[aa])