[BUG] Spec Advice - evaluate combinations takes too long

[jj-jimjams]

Describe the bug Spec advice looks at all possible combinations of the priorities added and evaluates which combination will be the most useful. This works fine when say three priorities are added. However, the longer the list, the list of possible combinations grows exponentially larger and it takes too long to run or times out.

To Reproduce Steps to reproduce the behavior:

1. at line 30 of main.py, enter: priorities_list = ("Land", "UpgradeBuild", "ExtraTile")
2. see how nice and quickly it runs (steps are timestamped) 3.at line 30 on main.py, enter: priorities_list = ("Land", "UpgradeBuild", "ExtraTile", "TileHonour", "FWMat") (and that's only 13 million combinations. If you add in "MaxQs" it goes to 25 million and repl.it kills it before it finishes.
3. go and have a cup of coffee and a nap as this will break the internet in your hour and takes 10 minutes to run

Expected behavior It all happens in seconds

*what i've explored so far

• sklearn - I took a stroll through this, but I don't know enough about statistics to really use this properly. It didn't seem to provide anything useful.
• scipy.optimize - I can't figure out how to get it to work well with non-metric values
• putting bounds on what is evaluations by number of nodes based on userSpecPoints. Encountered issues with the minimum band with high spec points and very few options selected
• a completely different method of getting what should be assigned by using one dataset (points, priorities) to predict the other dataset (points in each priority) (sklearn for this didn't provide anything useful but I only used a very small dataset and don't really know what I'm doing)

what might be useful

• some way of reducing the viable combinations (but I can't think of any)
• some way of getting it to guess a subset of combinations to try and eliminating others based on the result
• some other way of evaluating the best options
• magic

[jake5691]

well I certainly need some quite moments to understand what you did there...

Can you sketch the idea or correct my assumption:

1. you define all nodes like we did in DrawTest
2. have predefined groups which you use as some sort of priority list
3. go through those groups and create some sort of combinations of the nodes (what exactly you do there I don't understand) -> This uses the given groups but not the given spec points available, right? How do you determine if a node is activate able or not?
4. you use some machine learning algo to determine which combinations to go for

I haven't really checked all your code but it's kind of hard for me as I'm not much familiar with df, and you seem to always come back to those even when from my perspective some could be avoided and just done in the Spec-Class.

What I think might be a valid option to get this to work (not tested or anything):

1. define the nodes and give each node a usefulness score (maybe have 2 or 3 different sets of those (for speed, loyalty...) but give each node a score up front)
2. "just" run an optimisation algo to maximise the usefulness score with skill points as limitation. (I should know stuff about that from my studies but thats a while ago and mostly theoretically with some given cost functions, so not sure if I can get such an approach to work)

[jake5691]

forget what I wrote, I thought a little and my initial idea seams stupid or better, not feasible...

[jj-jimjams]

So that was my original idea. Assign a usefulness score to each node an then optimize to get the highest usefulness score with the constraint of userSpecPoints. I don't know how to write the algo for that so I'm essentially doing that in a bunch of code...

[jj-jimjams]

[jj-jimjams]

I have split it into two lists of three items each. Not a very elegant solution but it works for now