Sub-optimal Solutions in Uppaal-Stratego compared with Solutions Identified by Simulation

[samaye ismael]

Can Uppaal-Stratego generate a suboptimal solution compared to solutions identifiable through simulation? I have simulated 100 runs in an example, and I am finding values smaller than the minimum determined by Stratego.

[Marius Mikučionis]

On Tuesday, 3 October 2023 at 10:03:17 UTC+2 samayeismael wrote:

Can Uppaal-Stratego generate a suboptimal solution compared to solutions identifiable through simulation? I have simulated 100 runs in an example, and I am finding values smaller than the minimum determined by Stratego.

I guess you used the `minE` or `maxE` query which uses learning algorithm based on the sample runs it has seen.

Learning is not guaranteed to find the optimal solution. 

For example, if number of controllable options is large and the requested number of runs is small, then it is quite likely that it has not seen the best solution. 

Note that the default values for the number of runs are quite small:

This is optimized for quick feedback when developing models, but you can set it larger values and the quality of solutions should improve.

Best regards,

Marius

[samaye ismael]

Thank you for your response. In my case, I am using the minE function, and I would like to know which learning parameters I should change so that the tool takes into account all simulation cases. Thank you

cordialement SAMAYE Ismael

--
You received this message because you are subscribed to the Google Groups "UPPAAL" group.
To unsubscribe from this group and stop receiving emails from it, send an email to uppaal+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/uppaal/f3f832aa-b110-4285-ae64-45c98cf99ec0n%40googlegroups.com.

[Marius Mikučionis]

On Tuesday, 3 October 2023 at 20:13:50 UTC+2 samayeismael wrote:

Thank you for your response. In my case, I am using the minE function, and I would like to know which learning parameters I should change so that the tool takes into account all simulation cases. Thank you

Visiting all simulation cases is not always possible (and you probably do not want that as there can be exponentially many).

For example:

After 1 time unit we have 1 choice with 2 options -- overall 2 cases.

After 2 time units: 2 distinct choices -- overall 4 cases.

After 10 time units: 10 choices -- 1024 cases, need at least 1024 runs to visit and evaluate them all.

And so on, which is unsustainable. 

In such cases we estimate the number of cases and limit the number of choices by limiting the run duration in the query (also known as "online setup", where only a local strategy is synthesized at a time). 

In addition, if we can make an assumption that variable `i` has no influence over the result, then we can limit the visibility of the state to just location (and ignore the value of `i`), like:

strategy S = minE(cost)[<=limit] { P.location } -> {} : time>=limit

then Uppaal will treat all those choices as one and will not require that many runs. 

On the other hand, if variable `i` (or something else not mentioned in the visible part of state) does have influence on the result, then the learned strategy might not be as good, or maybe good enough. It is all about finding a good enough abstraction.

Best regards,
Marius

 

cordialement SAMAYE Ismael

[samaye ismael]

Hello, please, 

I would like to have your opinion on the results I obtain with Uppaal Stratego. Indeed, on the graph, we have in red the different possible values of parameter A, and in black its optimization. I would like to understand why the tool cannot provide a value lower than that of B, despite all the constraints I have defined. Thank you for your responses.