37 - Prob end and rewrite of the results #507
Conversation
Not terribly useful but I thought it made some of the tournament code I'm writing a tiny bit more readable.
Working through tests. Work also required on the results.
Am not sure what behaviour I actually want yet so commenting out. Need to work on the results set.
Every tournament is going to need the potential to have it's own Results class (not all current graphical outputs make sense for all potential tournament types).
Just to be able to loop over players if need be.
- normalised_scores_diff_length - normalised_payoff_diff_length - Also minor change to wording in docstring (both roots are real)
Replacing with players. This seems to be more in line with the rest of the library and obviously loses no functionality. I've chosen to make the internal attribute a list to be in line with tournament attribute (also players, also a list).
Result class directly calculates everything from Matches Init takes list of players and set of matches. No assumptions are made about the structures of the Matches passed. To do this, the Match class is made more powerful (it basically does all the work): - Adding a score method. - Adding a score per turn method to Match class. - Adding cooperation count to match class.
I believe there was a bug in the morality metric bug. It took count of plays vs self but other such metrics do not. This is easily fixable.
|
Thanks @marcharper, I've made those changes. For info in 627fe84 I've added a property based test that checks that |
|
Sorry, got a bit overtaken by incoming stuff today. It will have to be tomorrow now |
No problem at all, whenever you have time :) |
|
I'm starting to look at this and, overall, I think it looks good. One comment strikes me immediately:
If that's the case, then could you remove the redundant methods from those two modules? If either ends up empty as a result, then remove the module. I think the rest of my comments are likely to be picky stuff!! |
Yeah: good shout. On it! :) |
bea72af has done this (removed both the modules). |
| outcome (dict): returned from the Tournament class and containing | ||
| various sets of results for processing by this class. | ||
| matches (list): a list of list of completed matches | ||
| (1 for each repetition) | ||
| with_morality (bool): a flag to determine whether morality metrics | ||
| should be calculated. | ||
| """ |
There was a problem hiding this comment.
This whole docstring needs to be in numpy format as we agreed in #347
Whilst we've not bothered going through existing docstrings and changing them in one go for that ticket, we have been doing so whenever they've been 'touched' as part of other changes.
There was a problem hiding this comment.
Yup: good catch. I'm on it.
|
I think 796a4ae has got the numpy formatting but let me know if not. |
| self.good_partner_matrix, len(players), repetitions) | ||
| self.eigenjesus_rating = ac.eigenvector(self.normalised_cooperation) | ||
| self.eigenmoses_rating = ac.eigenvector(self.vengeful_cooperation) | ||
| self.matches = matches # List of dicts mapping tuples of players to matches |
There was a problem hiding this comment.
the comment here doesn't match with the docstring. I suggest it gets deleted and the docstring rules
There was a problem hiding this comment.
The docstring is actually incorrect, it's not a list of lists but a list of dictionaries. Fixing that now.
Also fixing docstring def for matches.
|
402ae74 gets all the numpy formatting. I've realised that the shorter methods in the |
|
Not worth the effort, I reckon |
|
I'm happy with this once travis passes. |
🎆 🎉 👍 🙇 |
You might get to return to the real world for a while |
This adds to #496 (which could be closed).
This is mainly a rewrite of the results module. Now it is initialised by being passed a set of interactions (list of dictionaries mapping indices to matches). This has the main advantage of simplifying things. No more
results.results,results.outputandresults.outcomewhich were various forms of holding the same data, now all data is basically inresults.matches.To do this:
payoff.py(which I think could be removed) and mostly moved everything out ofcooperation.py. I'm not sure I see the advantage of that extra layer but perhaps I'm missing something. All result calculations are done by manipulating matches and aim to make no assumptions about what matches are actually played.This second point has the advantage of making the probabilistic end tournament very simple and in fact I've gotten rid of the
ProbEndResultsclass as it is not at all needed.The
Resultsclass is now in essence a class that computes a standard set of results on any set of matches (even the tests examples I used in the class are not a standard round robin).One downside is that it is not taking advantage of the length of a match being constant for a standard tournament but I'd suggest that if need be it's easy to inherit and build a
Resultsclass that took advantage of the structure but I don't think that's at present necessary.Furthermore if at any point we wanted to add to the
standardset of results it's just a question of adding a method that only has one type of data source to deal with: the matches. I've kind of done an example of this: the results class now also has a match lengths attribute with just shows the lengths of matches played by each player (useless for standard tournament but neat to see expected results for prob end).Here's a notebook (which I will at some point tweak): https://github.com/Axelrod-Python/Axelrod-notebooks/blob/37/basic-prob-end-tournament.ipynb
Looking forward to hearing thoughts on this.
PS I actually changed one doctest in the morality metrics. I believe that was not quite being calculated consistently. Might be worth having a specific look at to make sure I'm correct. It's an easy fix back if I'm wrong.