Using aries-opt engine

[jdavidrcamacho]

Hello everyone,

I have a quick doubt about the engines used in unified-planning.
I know we can use the aries engine, but can we use the aries-opt engine (mentioned here)?

[arbimo]

Hi, aries-opt is not available by default because:

1. it is generally useless: there is very few cases where it would provide an optimality proof (specifically scheduling and non-recursive HTN). In all other cases, the solver would just run forever.
2. it would be misleading as the current problem kinds do not allow capturing these, specifically, missing the distinction between recursive and non recursive HTN.

If you are interested in plan quality, then you should use the AnytimePlanning operation mode to get a stream of solution of increasing quality. Just as aries-opt, it would stop with SOLVED_OPTIMALLY when the optimality of the last solution is proven.

If you really only want the provably optimal plan, you can register the up_aries.AriesOpt class manually: https://unified-planning.readthedocs.io/en/latest/notebooks/tutorial/planner-integration.html#Registering-the-engine

[jdavidrcamacho]

Thank you by the answer! I manage to register aries-opt and use it but my problem still remains maybe you can shed some light on why the reason might be or how to solve it.

This happens for both aries and aries-opt and sometimes when a plan is generated it gives a slightly different answer. For example while I would expect the optimal solution for my problem to be and most of the time is:

aries Hiearchical SequentialPlan:
    switch_on(instrument0, satellite0)
    nop
    turn_to(satellite0, groundstation2, phenomenon6)
    calibrate(satellite0, instrument0, groundstation2)
    turn_to(satellite0, phenomenon4, groundstation2)
    take_image(satellite0, phenomenon4, instrument0, thermograph0)
    nop
    turn_to(satellite0, star5, phenomenon4)
    take_image(satellite0, star5, instrument0, thermograph0)
    nop
    turn_to(satellite0, phenomenon6, star5)
    take_image(satellite0, phenomenon6, instrument0, thermograph0)

Sometimes it generates this

aries Hiearchical SequentialPlan:
    switch_on(instrument0, satellite0)
    nop
    turn_to(satellite0, groundstation2, phenomenon6)
    calibrate(satellite0, instrument0, groundstation2)
    turn_to(satellite0, phenomenon4, groundstation2)
    take_image(satellite0, phenomenon4, instrument0, thermograph0)
    switch_off(instrument0, satellite0)
    switch_on(instrument0, satellite0)
    nop
    turn_to(satellite0, groundstation2, phenomenon4)
    calibrate(satellite0, instrument0, groundstation2)
    turn_to(satellite0, star5, groundstation2)
    take_image(satellite0, star5, instrument0, thermograph0)
    nop
    turn_to(satellite0, phenomenon6, star5)
    take_image(satellite0, phenomenon6, instrument0, thermograph0)

Where it has steps that are unnecessary steps, i.e. switching a satellite on and off for no apparent reason. Do you know how can I make aries results to be deterministic and truly optimal?

[arbimo]

The solver will run a handful of strategies in parallel that would share solutions and no-good clauses, hence the search is not expected to be deterministic.

What is the metric you specified for the problem ? If there is no metric, the solver may return any plan. If you specifically require length-optimal plan aries-opt shouldn't and I would be interested in seeing the problem to analyze it.

[jdavidrcamacho]

I have tried 3 different ways:

OneshotPlanner(name="aries", optimality_guarantee=PlanGenerationResultStatus.SOLVED_OPTIMALLY,),

OneshotPlanner(name="aries", optimality_guarantee=PlanGenerationResultStatus.SOLVED_SATISFICING,), and

AnytimePlanner(name="aries", anytime_guarantee="INCREASING_QUALITY")

where I also used problem.add_quality_metric(up.model.metrics.MinimizeSequentialPlanLength())

but still the weird plan show up from time to time (~ every 10 times I run it, 9 succeeded but 1 has those unnecessary steps)

[arbimo]

Well, you always use aries in the above examples. Try one of the following.

solver = OneshotPlanner(name="aries-opt")
# or directly
sovler = up_aries.AriesOpt()

As you specify the name "aries", this is the planner that would be selected. Also your optimality_guarantee parameter has the wrong type.

[jdavidrcamacho]

I didn't mentioned but already tried OneshotPlanner(name="aries-opt") and after a few trials it gave the same wrong results

[arbimo]

Then, could please send what's necessary to reproduce the problem.

[jdavidrcamacho]

Can you please send me an email ([email protected])? I'll reply to it with the data and scripts

[jdavidrcamacho]

Closing it as we discussed the issue via email

[arbimo]

FTR, the problem was that to have an optimal plan, both an optimal planner and a metric is needed.