Question on Using StormPy to Extract Scheduler for MDP Model

[Wendybeefish]

Hello,

I have an MDP model where I need the scheduler to generate the optimal policy while maximizing rewards. I have a PRISM file with modules for transitions and for rewards (module "all_rewards").

When I run this .pm file and property in PRISM and in Storm CLI, both run smoothly and produce the same policy. However, when I run it in StormPy, I get a totally wrong set of policies.... It seems like the model is not reading my rewards correctly or is not actually maximizing them. I am not sure which setting I might have gotten wrong. Below is my current StormPy script:

import stormpy

Input

prism_file = "5_updated.pm" formula_str = 'R{"all_rewards"}max=? [ F "termination" ]'

Parse model and property

prism_program = stormpy.parse_prism_program(prism_file) properties = stormpy.parse_properties(formula_str, prism_program)

Builder options

options = stormpy.BuilderOptions() options.set_build_state_valuations(True) options.set_build_choice_labels(True) options.set_build_with_choice_origins(True)

Build and solve

model = stormpy.build_sparse_model_with_options(prism_program, options) result = stormpy.model_checking(model, properties[0], extract_scheduler=True) scheduler = result.scheduler

Export policy

for state in model.states: choice = scheduler.get_choice(state) action_index = choice.get_deterministic_choice() action = state.actions[action_index] f"state {state.id}: {state.valuations} => {action.labels}\n"

Could you please let me know if there is any setting I might be missing that ensures the rewards are correctly read and applied during scheduler synthesis, and if I am following the correct steps to extract the scheduler for this type of reward-maximization property?

Thank you for your time and guidance.

Best regards, Wendy

[volkm]

On a first look, your script seems okay for me and is very much the same as the example. Can you share your model such that we can better investigate the issue? If you do not want to publicly share the model, you can send us the file via email at [email protected].