Commit df7f9e55 authored by Steven Cordwell's avatar Steven Cordwell
Browse files

fix some typos in docstrings

parent 118a2fa8
...@@ -70,46 +70,50 @@ class MDP(object): ...@@ -70,46 +70,50 @@ class MDP(object):
"""A Markov Decision Problem. """A Markov Decision Problem.
Let S = the number of states, and A = the number of acions. Let ``S`` = the number of states, and ``A`` = the number of acions.
Parameters Parameters
---------- ----------
transitions : array transitions : array
Transition probability matrices. These can be defined in a variety of Transition probability matrices. These can be defined in a variety of
ways. The simplest is a numpy array that has the shape (A, S, S), ways. The simplest is a numpy array that has the shape ``(A, S, S)``,
though there are other possibilities. It can be a tuple or list or though there are other possibilities. It can be a tuple or list or
numpy object array of length A, where each element contains a numpy numpy object array of length ``A``, where each element contains a numpy
array or matrix that has the shape (S, S). This "list of matrices" form array or matrix that has the shape ``(S, S)``. This "list of matrices"
is useful when the transition matrices are sparse as form is useful when the transition matrices are sparse as
scipy.sparse.csr_matrix matrices can be used. In summary, each action's ``scipy.sparse.csr_matrix`` matrices can be used. In summary, each
transition matrix must be indexable like ``P[a]`` where action's transition matrix must be indexable like ``transitions[a]``
``a`` ∈ {0, 1...A-1}. where ``a`` ∈ {0, 1...A-1}, and ``transitions[a]`` returns an ``S`` ×
``S`` array-like object.
reward : array reward : array
Reward matrices or vectors. Like the transition matrices, these can Reward matrices or vectors. Like the transition matrices, these can
also be defined in a variety of ways. Again the simplest is a numpy also be defined in a variety of ways. Again the simplest is a numpy
array that has the shape (S, A), (S,) or (A, S, S). A list of lists can array that has the shape ``(S, A)``, ``(S,)`` or ``(A, S, S)``. A list
be used, where each inner list has length S. A list of numpy arrays is of lists can be used, where each inner list has length ``S`` and the
possible where each inner array can be of the shape (S,), (S, 1), outer list has length ``A``. A list of numpy arrays is possible where
(1, S) or (S, S). Also scipy.sparse.csr_matrix can be used instead of each inner array can be of the shape ``(S,)``, ``(S, 1)``, ``(1, S)``
numpy arrays. In addition, the outer list can be replaced with a tuple or ``(S, S)``. Also ``scipy.sparse.csr_matrix`` can be used instead of
or numpy object array can be used. numpy arrays. In addition, the outer list can be replaced by any object
that can be indexed like ``reward[a]`` such as a tuple or numpy object
array of length ``A``.
discount : float discount : float
Discount factor. The per time-step discount factor on future rewards. Discount factor. The per time-step discount factor on future rewards.
Valid values are greater than 0 upto and including 1. If the discount Valid values are greater than 0 upto and including 1. If the discount
factor is 1, then convergence is cannot be assumed and a warning will factor is 1, then convergence is cannot be assumed and a warning will
be displayed. Subclasses of ``MDP`` may pass None in the case where the be displayed. Subclasses of ``MDP`` may pass ``None`` in the case where
algorithm does not use a discount factor. the algorithm does not use a discount factor.
epsilon : float epsilon : float
Stopping criterion. The maximum change in the value function at each Stopping criterion. The maximum change in the value function at each
iteration is compared against ``epsilon``. Once the change falls below iteration is compared against ``epsilon``. Once the change falls below
this value, then the value function is considered to have converged to this value, then the value function is considered to have converged to
the optimal value function. Subclasses of ``MDP`` may pass None in the the optimal value function. Subclasses of ``MDP`` may pass ``None`` in
case where the algorithm does not use a stopping criterion. the case where the algorithm does not use an epsilon-optimal stopping
criterion.
max_iter : int max_iter : int
Maximum number of iterations. The algorithm will be terminated once Maximum number of iterations. The algorithm will be terminated once
this many iterations have elapsed. This must be greater than 0 if this many iterations have elapsed. This must be greater than 0 if
specified. Subclasses of ``MDP`` may pass None in the case where the specified. Subclasses of ``MDP`` may pass ``None`` in the case where
algorithm does not use a maximum number of iterations. the algorithm does not use a maximum number of iterations.
Attributes Attributes
---------- ----------
...@@ -130,12 +134,12 @@ class MDP(object): ...@@ -130,12 +134,12 @@ class MDP(object):
time : float time : float
The time used to converge to the optimal policy. The time used to converge to the optimal policy.
verbose : boolean verbose : boolean
Whether verbose output should be displayed in not. Whether verbose output should be displayed or not.
Methods Methods
------- -------
run run
Implemented in child classes as the main algorithm loop. Raises and Implemented in child classes as the main algorithm loop. Raises an
exception if it has not been overridden. exception if it has not been overridden.
setSilent setSilent
Turn the verbosity off Turn the verbosity off
...@@ -314,11 +318,11 @@ class FiniteHorizon(MDP): ...@@ -314,11 +318,11 @@ class FiniteHorizon(MDP):
--------------- ---------------
V : array V : array
Optimal value function. Shape = (S, N+1). ``V[:, n]`` = optimal value Optimal value function. Shape = (S, N+1). ``V[:, n]`` = optimal value
function at stage ``n`` with stage in (0, 1...N-1). ``V[:, N]`` value function at stage ``n`` with stage in {0, 1...N-1}. ``V[:, N]`` value
function for terminal stage. function for terminal stage.
policy : array policy : array
Optimal policy. ``policy[:, n]`` = optimal policy at stage ``n`` with Optimal policy. ``policy[:, n]`` = optimal policy at stage ``n`` with
stage in (0, 1...N). ``policy[:, N]`` = policy for stage ``N``. stage in {0, 1...N}. ``policy[:, N]`` = policy for stage ``N``.
time : float time : float
used CPU time used CPU time
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment