Adaptive Zero Determinant Strategy

[dashiellfryer]

My recent graduate Emmanuel Estrada and I made a class that adapts its four vector in response to payoff. The player maintains zero determinant bounds each round. In our tests this player eventually outperforms all the standard zero determinant strategies. We tested by using 2000 rounds instead of the standard 200. We look forward to any comments.

[drvinceknight]

Thanks a lot for contributing this!

The tests are not passing but I believe this is not due to your code contribution but to an upgrade of pandas which I've captured here #1283. I don't suggest you fix that here as it involves the result set which can be a bit of a pain so for now could you modify one line in the requirements.txt file. Line 7 currently reads:

pandas>=0.18.1

Could you change that to be:

pandas==0.25.3

This will just make sure the test suite runs correctly and we'll fix the pandas issue elsewhere.

I haven't looked at the logic itself yet but we are going to need some automated tests for the strategy as well. Is this something you would like a hand with?

I've also left some very minor requests for formatting/style changes.

Thanks again for this contribution, I look forward to getting it in :)

[drvinceknight]

This is looking good, a few requests in terms of writing the tests and adding some documentation. Happy to help with that.

[drvinceknight]

Could we add few tests for the two specific methods: _adjust_parameters and score_last_round (just running the methods and checking that the attributes have changed accordingly?

[drvinceknight]

This tests need to be tweaked so that we can check expected behaviours. There's information on writing tests here: https://axelrod.readthedocs.io/en/stable/tutorials/contributing/strategy/writing_test_for_the_new_strategy.html

For example this specific paragraph would be something like (although I am not sure what the expected behaviour is):

Suggested change

	player = self.player(l=R)
	axelrod.seed(0)
	match = axelrod.Match((player, axelrod.Alternator()), turns=200)
	match.play()
	expected_actions = [(C,D), (D, C)] * 100
	self.versus_test(opponent=axelrod.Alternator),
	expected_actions=expected_actions, seed=0, init_kwargs={"l": R})

Note that we could also check that the other attributes have changed accordingly using the attrs keyword.

Let me know if I can assist with that (happy to PR the tests to your branch).

[marcharper]

I agree but I didn't want to search for seeds just to have to do so again once #1288 goes in. The current tests make sure that at least each branch is run (so that the limits are bounced into).

[drvinceknight]

I understand. Are you suggesting this are place holders and we merge #1288 and then fix these?

[marcharper]

Yes, after we merge this I'll update all the tests that require seeds in #1288 (there are many and most will have to change).

[drvinceknight]

Sounds good. Happy to take a share of that work if helpful.

[drvinceknight]

It would be helpful to have some more information here, perhaps an explanation of the update procedure? It's ok to go quite in depth and include some mathematics which we can then use to check that the _adjust_parameters method is working as expected.

[marcharper]

Hi all, quick update on this one. There are really four possibly strategies here depending on how one wants to update the underlying parameters each round. It's also not clear how much to update each parameter, so I'd like to run some evolutionary algorithms to try to fit these hyperparameters.