Feature step dependencies are a pain to work with about aslib-spec

[berndbischl]

i guees i agrree. how do we parse yaml in r?

[larskotthoff]

https://cran.r-project.org/web/packages/yaml/index.html

[berndbischl]

ok sorry :)

[larskotthoff]

No worries :)
As far as I can see at the moment this would give us exactly the same data structure as we have at the moment (apart from the feature steps of course), so any changes should be minimal.

[mlindauer]

Hi Lars,

I agree that the current format of the feature groups is an issue.
I also like the idea of "provide" and "requires".

However, please note first of all,
that your example is wrong.
Pre provides the following features: "reducedVars,nvars,nclausesOrig,nvarsOrig,nclauses,reducedClauses".

Furthermore, I don't like YAML so much.
We use it for one of our homepages and it is always a pain to edit the yaml files.
Aren't there any better alternatives?

Cheers,
Marius

[larskotthoff]

Ok, that the example is wrong would have been much clearer in the new format :)

I don't see how editing YAML is more painful than editing a non-standard format.

[mlindauer]

In the end, I can live with YAML.
however, there is no way to specify this with arff, or?
If possible, I would like to prevent to use two different standard formats.

[larskotthoff]

Again, I don't see how using two different standard formats is worse than using a standard and a non-standard format. In principle I don't have a problem with using YAML for everything.

[mlindauer]

How would one of the other files look like in YAML?
I read in wikipedia that each JSON file is also valid YAML (>=1.2) file.
I like JSON but I don't know whether this is really user-friendly.

[larskotthoff]

Hmm, I guess something like

- instance_id: bla
  repetition: 1
  feature_x: foo

I don't really see a problem with being user friendly -- you're not supposed to edit/write those files manually.

[mlindauer]

such a format would blow up our files by more than factor 2 I guess.

The description.txt is a file I always write manually.

[berndbischl]

you can forget arff for such files immediatly

[larskotthoff]

Yes, everything would be much larger. But as I said, I'm not opposed to keeping everything but description.txt in arff. We also have citation.bib which is in yet another standard format.

[mlindauer]

OK.

I also asked Matthias whether he likes this new format. and he agreed.
so, please go on and make the changes.

Cheers,
Marius

[larskotthoff]

Ok, what's your feeling on making the lists proper YAML lists as well? I.e. instead of comma-separated they would be

provides:
  - CG_mean
  - CG_coeff_variation
  - etc.

[mlindauer]

I like the comma-separated more since I can look up the corresponding feature step to a feature by looking one line above (and not n lines).
To have a proper YAML (1.2), which is similiar to right now, we could use

[CG_mean, CG_coeff_variation,...]

However, we should change the entire file.
So for example also algorithms_deterministic.

[larskotthoff]

Ok, but presumably you're not going to parse the YAML yourself but use a library? And yes, that would apply for everything -- if the data structure is serialized by a YAML library we may not even be able to control which type of list we get (and don't need to care).

So I guess my real question is whether you're planning to use a library to parse/write the YAML.

[berndbischl]

parsing: of course.

but i would prefer it if people could still manually write (smaller) files without programing.

can we do that?

[mlindauer]

I often have a look into the description.txt files to get a better feeling for the scenarios, e.g., which algorithms are used; how many feature are used and how are the feature distributed in the feature groups.
I could write scripts for such things, but looking into the files is often faster.
So I would prefer that I can easily read the files.

[berndbischl]

well that argument i find slightly strange? why not use the eda overview?

[larskotthoff]

Of course you can still read/write the files manually and that shouldn't even be much more difficult than it is now. But it would be much easier to parse/write programmatically because we can just use YAML libraries.

[berndbischl]

i meam we invested lots of time to write exactly scripts for that purpose.... web based.....

[larskotthoff]

Which, come to think of it, we should rerun to update the web pages at some point.

[berndbischl]

Proposal: Use travis for that. People do PRs for a new scenario. Then travis builds all EDA stuff. This even checks the validity of the scenario files. Only then we merge. The only thing we then have to run manually might be the selector benchmarks.

[larskotthoff]

+1

[mlindauer]

i meam we invested lots of time to write exactly scripts for that purpose.... web based.....

1. I'm not always online.
2. I'm faster with my local files than finding the URL and then clicking through the web interface.

[larskotthoff]

Ok, so you think that

- name: Basic
  provides:
    - vars_clauses_ratio
    - POSNEG_RATIO_CLAUSE_mean
    - POSNEG_RATIO_CLAUSE_coeff_variation
    - POSNEG_RATIO_CLAUSE_min
    - POSNEG_RATIO_CLAUSE_max
    - POSNEG_RATIO_CLAUSE_entropy
    - VCG_CLAUSE_mean
    - VCG_CLAUSE_coeff_variation
    - VCG_CLAUSE_min
    - VCG_CLAUSE_max
    - VCG_CLAUSE_entropy
    - UNARY
    - BINARYp
    - TRINARYp
  requires: Pre

is harder to read than

- name: Basic
  provides: vars_clauses_ratio,POSNEG_RATIO_CLAUSE_mean,POSNEG_RATIO_CLAUSE_coeff_variation,POSNEG_RATIO_CLAUSE_min,POSNEG_RATIO_CLAUSE_max,POSNEG_RATIO_CLAUSE_entropy,VCG_CLAUSE_mean,VCG_CLAUSE_coeff_variation,VCG_CLAUSE_min,VCG_CLAUSE_max,VCG_CLAUSE_entropy,UNARY,BINARYp,TRINARYp
  requires: Pre

[mlindauer]

Yes, but in the end, I don't feel strongly about this.
So, I can also live with the first format if we don't have a nice way to automatically generate the second format.

[larskotthoff]

Ok, I've updated the spec, converted all the scenarios and updated the R code.

@mlindauer Could you please update the Python code/checker?

[mlindauer]

I'm on vacation for the next two weeks. I will do it afterwards.

[larskotthoff]

Ok, thanks. No rush :)

[larskotthoff]

It just occurred to me that we should also have a look at the feature_runstatus.arff files for instances that are presolved. The spec doesn't say what should happen to dependent feature steps in this case and the data is inconsistent. For example for ASP, feature steps that depend on one that presolved seem to be listed as presolved" as well but the costs aren't given, implying that they weren't actually run. For the SAT data sets, the runstatus of feature steps that depend on one that presolved are listed as unknown (which probably makes more sense in this case).

[mlindauer]

Hi,

I started to implement the new description.txt parser and I found an issue.
According to the spec, "performance_measures" specifies a list.
But looking at some of the description.txt files, e.g., ASP-POTASSCO, it is only a string:
performance_measures: runtime

So, the format according to YAML should be:
performance_measures:
- runtime

The same issue holds for "maximize" and "performance_type".

[mlindauer]

The same issue applies to feature_step->"requires" in same senarios.
In ASP-POTASSCO it is fine:

    Dynamic-1:
        requires:
            - Static

IN SAT11-HAND it is not OK:

    Basic:
      requires: Pre

[mlindauer]

I updated the checker tool (and flexfolio).
Right now, the checker tools complains about the issues raised above.

[larskotthoff]

Thanks, good catch. Could you fix the files please?

[mlindauer]

Hi, I fixed it. All scenarios in the master branch are now compatible with the checker tool again.

However, I found another issue.
At some point, we agreed that we need an order of the feature steps. This was implicitly given by the order of the feature steps in the description.txt.
Since, we use YAML now, we encode the "feature_steps" as dictionaries:

feature_steps:
    Pre:
      provides:
        - nvarsOrig
       [...]
    Basic:
      requires: 
        - Pre

Parsing this file (at least with Python) will give you a dictionary without a defined order of the feature steps. So, we either have to change "feature_steps" to list (which would look unintuitively and ugly imho) or we add another list, such as "feature_step_order".
What do you think?

Cheers,
Marius

[larskotthoff]

Just remind me what the order is needed for? You can derive any ordering constraints from the provides/requires right?

[mlindauer]

If I correctly remember, the problem was the presolved feature steps.

1. The features were computed in an (unknown) order and if a feature step pre-solved an instance, the remaining feature steps were not computed (at least true for ASP and SAT scenarios).
2. We discussed the definition of the oracle at some point. If we want to include the feature steps as a possible algorithm to solve an instance (important for some scenarios) in the oracle defintion, we have to know the right order of the feature steps, or else we have to solve an NP-hard problem (i.e., all possible orders of feature steps) to find an optimal order.
   (3. there is more than one possible order, if we only consider "requires")

[larskotthoff]

1. Sounds to me like we should have a feature runstatus "not computed" then -- using the order to derive this is quite similar to how the dependencies were encoded. Not at all obvious and intuitive and bound to trip somebody up.
2. I remember -- what conclusion did we come to? It seems fair enough to me that oracle would be able to change the ordering of feature steps.
3. I don't see that as a disadvantage.

[mlindauer]

1. Saying some features are "not computed" sounds even more unintuitive. Without an order of the feature steps, it is not explained why they are not computed. And we assume so far that the data is complete as long as there are no good reasons for missing data.
2. I think we postponed the discussion to later and used simply the old definition of the oracle without consideration of feature steps.

[larskotthoff]

Ok, so let's have a feature status "not computed because instance presolved by previous feature step". We don't need to know what that feature step was, do we?

[mlindauer]

OK, I agree that we should have something like "not computed because instance presolved by previous feature step".
However, if we have such a status, I still think we should have some more information about the order of the feature steps - at least how they were generated; the user can still decide to use another order.
The arguments for such information are:

1. We would know which step was responsible for this new status
2. The optimal order of the feature steps (-> presolved status) is exactly the order in which the features were generated. I don't see an argument why the users should try to figure this out by themselves if we already know it. (In the same way, it is also important for a new oracle definition as mentioned before.)

[larskotthoff]

Should the order of the feature steps used when generating the data for the scenarios be part of the metadata?

[mlindauer]

Yes?

[larskotthoff]

Ok, then let's do that.