Sebastian Junges
6 years ago
6 changed files with 305 additions and 0 deletions
-
1doc/source/advanced_topics.rst
-
106doc/source/doc/exploration.rst
-
30examples/exploration/01-exploration.py
-
38examples/exploration/02-exploration.py
-
4lib/stormpy/examples/files.py
-
126lib/stormpy/examples/files/mdp/maze_2.nm
@ -0,0 +1,106 @@ |
|||
**************** |
|||
Exploring Models |
|||
**************** |
|||
|
|||
Background |
|||
===================== |
|||
|
|||
Often, stormpy is used as a testbed for new algorithms. |
|||
An essential step is to transfer the (low-level) descriptions of an MDP or other state-based model into |
|||
an own algorithm. In this section, we discuss some of the functionality. |
|||
|
|||
Reading MDPs |
|||
===================== |
|||
|
|||
.. seealso:: `01-exploration.py <https://github.com/moves-rwth/stormpy/blob/master/examples/exploration/01-exploration.py>`_ |
|||
|
|||
In :doc:`../getting_started`, we briefly iterated over a DTMC. In this section, we explore an MDP:: |
|||
|
|||
>>> import doctest |
|||
>>> doctest.ELLIPSIS_MARKER = '-etc-' # doctest:+ELLIPSIS |
|||
>>> import stormpy |
|||
>>> import stormpy.examples |
|||
>>> import stormpy.examples.files |
|||
>>> program = stormpy.parse_prism_program(stormpy.examples.files.prism_mdp_maze) |
|||
>>> prop = "R=? [F \"goal\"]" |
|||
|
|||
>>> properties = stormpy.parse_properties_for_prism_program(prop, program, None) |
|||
>>> model = stormpy.build_model(program, properties) |
|||
|
|||
The iteration over the model is as before, but now, for every action, we can have several transitions:: |
|||
|
|||
>>> for state in model.states: |
|||
... if state.id in model.initial_states: |
|||
... print("State {} is initial".format(state.id)) |
|||
... for action in state.actions: |
|||
... for transition in action.transitions: |
|||
... print("From state {} by action {}, with probability {}, go to state {}".format(state, action, transition.value(), transition.column)) |
|||
-etc- |
|||
|
|||
The output (omitted for brievety) contains sentences like: |
|||
|
|||
From state 1 by action 0, with probability 1.0, go to state 2 |
|||
From state 1 by action 1, with probability 1.0, go to state 1 |
|||
|
|||
|
|||
|
|||
Internally, storm can hold hints to the origin of the actions, which may be helpful to give meaning and for debugging. |
|||
As the availability and the encoding of this data depends on the input model, we discuss these features in :doc:`highlevel_models`. |
|||
|
|||
|
|||
Storm currently supports deterministic rewards on states or actions. More information can be found in that :doc:`reward_models`. |
|||
|
|||
|
|||
Reading POMDPs |
|||
====================== |
|||
.. seealso:: `02-exploration.py <https://github.com/moves-rwth/stormpy/blob/master/examples/exploration/01-exploration.py>`_ |
|||
|
|||
|
|||
Internally, POMDPs extend MDPs. Thus, iterating over the MDP is done as before. |
|||
|
|||
>>> import stormpy |
|||
>>> import stormpy.examples |
|||
>>> import stormpy.examples.files |
|||
>>> program = stormpy.parse_prism_program(stormpy.examples.files.prism_pomdp_maze) |
|||
>>> prop = "R=? [F \"goal\"]" |
|||
|
|||
>>> properties = stormpy.parse_properties_for_prism_program(prop, program, None) |
|||
>>> model = stormpy.build_model(program, properties) |
|||
|
|||
Indeed, all that changed in the code above is the example we use. |
|||
And, that the model type now is a POMDP:: |
|||
|
|||
>>> print(model.model_type) |
|||
ModelType.POMDP |
|||
|
|||
Additionally, POMDPs have a set of observations, which are internally just numbered by an integer from 0 to the number of observations -1 :: |
|||
|
|||
>>> print(model.nr_observations) |
|||
8 |
|||
>>> for state in model.states: |
|||
... print("State {} has observation id {}".format(state.id, model.observations[state.id])) |
|||
State 0 has observation id 6 |
|||
State 1 has observation id 1 |
|||
State 2 has observation id 4 |
|||
State 3 has observation id 7 |
|||
State 4 has observation id 4 |
|||
State 5 has observation id 3 |
|||
State 6 has observation id 0 |
|||
State 7 has observation id 0 |
|||
State 8 has observation id 0 |
|||
State 9 has observation id 0 |
|||
State 10 has observation id 0 |
|||
State 11 has observation id 0 |
|||
State 12 has observation id 2 |
|||
State 13 has observation id 2 |
|||
State 14 has observation id 4 |
|||
State 15 has observation id 5 |
|||
|
|||
|
|||
|
|||
|
|||
|
|||
Reading MAs |
|||
====================== |
|||
|
|||
To be continued... |
@ -0,0 +1,30 @@ |
|||
import stormpy |
|||
import stormpy.core |
|||
|
|||
import stormpy.examples |
|||
import stormpy.examples.files |
|||
|
|||
|
|||
def example_exploration_01(): |
|||
""" |
|||
Example to exploration of MDPs. |
|||
:return: |
|||
""" |
|||
program = stormpy.parse_prism_program(stormpy.examples.files.prism_pomdp_maze) |
|||
prop = "R=? [F \"goal\"]" |
|||
properties = stormpy.parse_properties_for_prism_program(prop, program, None) |
|||
model = stormpy.build_model(program, properties) |
|||
print(model.model_type) |
|||
|
|||
|
|||
for state in model.states: |
|||
if state.id in model.initial_states: |
|||
print(state) |
|||
for action in state.actions: |
|||
for transition in action.transitions: |
|||
print("From state {} by action {}, with probability {}, go to state {}".format(state, action, transition.value(), |
|||
transition.column)) |
|||
|
|||
|
|||
if __name__ == '__main__': |
|||
example_exploration_01() |
@ -0,0 +1,38 @@ |
|||
import stormpy |
|||
import stormpy.core |
|||
|
|||
import stormpy.examples |
|||
import stormpy.examples.files |
|||
|
|||
|
|||
def example_exploration_02(): |
|||
""" |
|||
Example to exploration of POMDPs. |
|||
:return: |
|||
""" |
|||
program = stormpy.parse_prism_program(stormpy.examples.files.prism_pomdp_maze) |
|||
prop = "R=? [F \"goal\"]" |
|||
properties = stormpy.parse_properties_for_prism_program(prop, program, None) |
|||
model = stormpy.build_model(program, properties) |
|||
print(model.model_type) |
|||
# Internally, POMDPs are just MDPs with additional observation information. |
|||
# Thus, data structure exploration for MDPs can be applied as before. |
|||
initial_state = model.initial_states[0] |
|||
|
|||
for state in model.states: |
|||
if state.id in model.initial_states: |
|||
print(state) |
|||
for action in state.actions: |
|||
for transition in action.transitions: |
|||
print("From state {} by action {}, with probability {}, go to state {}".format(state, action, transition.value(), |
|||
transition.column)) |
|||
|
|||
print(model.nr_observations) |
|||
|
|||
for state in model.states: |
|||
print("State {} has observation id {}".format(state.id, model.observations[state.id])) |
|||
|
|||
|
|||
|
|||
if __name__ == '__main__': |
|||
example_exploration_02() |
@ -0,0 +1,126 @@ |
|||
|
|||
|
|||
// maze example (POMDP) |
|||
// slightly extends that presented in |
|||
// Littman, Cassandra and Kaelbling |
|||
// Learning policies for partially observable environments: Scaling up |
|||
// Technical Report CS, Brown University |
|||
// gxn 29/01/16 |
|||
// Made into a MDP for documentation of stormpy. |
|||
|
|||
// state space (value of variable "s") |
|||
|
|||
// 0 1 2 3 4 |
|||
// 5 6 7 |
|||
// 8 9 10 |
|||
// 11 13 12 |
|||
|
|||
// 13 is the target |
|||
|
|||
mdp |
|||
|
|||
|
|||
module maze |
|||
|
|||
s : [-1..13]; |
|||
|
|||
// initialisation |
|||
[] s=-1 -> 1/13 : (s'=0) |
|||
+ 1/13 : (s'=1) |
|||
+ 1/13 : (s'=2) |
|||
+ 1/13 : (s'=3) |
|||
+ 1/13 : (s'=4) |
|||
+ 1/13 : (s'=5) |
|||
+ 1/13 : (s'=6) |
|||
+ 1/13 : (s'=7) |
|||
+ 1/13 : (s'=8) |
|||
+ 1/13 : (s'=9) |
|||
+ 1/13 : (s'=10) |
|||
+ 1/13 : (s'=11) |
|||
+ 1/13 : (s'=12); |
|||
|
|||
// moving around the maze |
|||
|
|||
[east] s=0 -> (s'=1); |
|||
[west] s=0 -> (s'=0); |
|||
[north] s=0 -> (s'=0); |
|||
[south] s=0 -> (s'=5); |
|||
|
|||
[east] s=1 -> (s'=2); |
|||
[west] s=1 -> (s'=0); |
|||
[north] s=1 -> (s'=1); |
|||
[south] s=1 -> (s'=1); |
|||
|
|||
[east] s=2 -> (s'=3); |
|||
[west] s=2 -> (s'=1); |
|||
[north] s=2 -> (s'=2); |
|||
[south] s=2 -> (s'=6); |
|||
|
|||
[east] s=3 -> (s'=4); |
|||
[west] s=3 -> (s'=2); |
|||
[north] s=3 -> (s'=3); |
|||
[south] s=3 -> (s'=3); |
|||
|
|||
[east] s=4 -> (s'=4); |
|||
[west] s=4 -> (s'=3); |
|||
[north] s=4 -> (s'=4); |
|||
[south] s=4 -> (s'=7); |
|||
|
|||
[east] s=5 -> (s'=5); |
|||
[west] s=5 -> (s'=5); |
|||
[north] s=5 -> (s'=0); |
|||
[south] s=5 -> (s'=8); |
|||
|
|||
[east] s=6 -> (s'=6); |
|||
[west] s=6 -> (s'=6); |
|||
[north] s=6 -> (s'=2); |
|||
[south] s=6 -> (s'=9); |
|||
|
|||
[east] s=7 -> (s'=7); |
|||
[west] s=7 -> (s'=7); |
|||
[north] s=7 -> (s'=4); |
|||
[south] s=7 -> (s'=10); |
|||
|
|||
[east] s=8 -> (s'=8); |
|||
[west] s=8 -> (s'=8); |
|||
[north] s=8 -> (s'=5); |
|||
[south] s=8 -> (s'=11); |
|||
|
|||
[east] s=9 -> (s'=9); |
|||
[west] s=9 -> (s'=9); |
|||
[north] s=9 -> (s'=6); |
|||
[south] s=9 -> (s'=13); |
|||
|
|||
[east] s=10 -> (s'=9); |
|||
[west] s=10 -> (s'=9); |
|||
[north] s=10 -> (s'=7); |
|||
[south] s=10 -> (s'=12); |
|||
[east] s=11 -> (s'=11); |
|||
[west] s=11 -> (s'=11); |
|||
[north] s=11 -> (s'=8); |
|||
[south] s=11 -> (s'=11); |
|||
|
|||
[east] s=12 -> (s'=12); |
|||
[west] s=12 -> (s'=12); |
|||
[north] s=12 -> (s'=10); |
|||
[south] s=12 -> (s'=12); |
|||
|
|||
// loop when we reach the target |
|||
[done] s=13 -> true; |
|||
|
|||
endmodule |
|||
|
|||
// reward structure (number of steps to reach the target) |
|||
rewards |
|||
|
|||
[east] true : 1; |
|||
[west] true : 1; |
|||
[north] true : 1; |
|||
[south] true : 1; |
|||
|
|||
endrewards |
|||
|
|||
// target observation |
|||
label "goal" = s=13; |
|||
|
|||
|
Write
Preview
Loading…
Cancel
Save
Reference in new issue