Implemented Harrington (k75r from Axelrod's Second) #1146

gaffney2010 · 2017-12-04T13:56:36Z

Below are the fingerprints for the basic strategies, the random, and against the six custom strategies that I used in the test file.

I've also attached a PDF that bridges the gap from the Fortran to what I have, along with some small amount of commentary on the code.

HarringtonNotes.pdf

drvinceknight

This looks like a big effort! As far as I can tell everything looks good.

drvinceknight · 2017-12-04T15:23:15Z

axelrod/strategies/axelrod_second.py

+    1. Check for defect and parity streaks.
+    2. Check if cooperations are scheduled.
+    3. Otherwise,
+       - If turn < 37, Tit-for-Tat.


Need a blank line here (this is what's failing on the CI):

3. Otherwise, - If turn ...

drvinceknight · 2017-12-04T15:23:50Z

axelrod/strategies/axelrod_second.py

+
+class Harrington(Player):
+    """
+    Strategy submitted to Axelrod's second tournament by Paul Harringtn (K75R),


Harrington

drvinceknight · 2017-12-04T15:24:04Z

axelrod/strategies/axelrod_second.py

+    Strategy submitted to Axelrod's second tournament by Paul Harringtn (K75R),
+    and came in eighth in that tournament.
+
+    This strategy has three modes:  Normal, and Fair-weather, Defect.


Normal, Fair-weather and Defect

Looking at the source code https://github.com/Axelrod-Python/TourExec/blob/master/src/strategies/K75R.f, these names are of your choosing? Nothing wrong with that. Perhaps, we can add a sentence saying that though (just to avoid potential confusion):

..., and Defect. These mode names were not present in Harrington's submission.

Could we also add a sentence along the lines of what you wrote in the pdf: that the Fair-weather mode occurs only in 1 very specific type of situation.

drvinceknight · 2017-12-04T15:46:16Z

axelrod/strategies/axelrod_second.py

+
+        if turn < 37:
+            return self.try_return(opponent.history[-1], inc_parity=True)
+        elif turn == 37:


This can just be if turn == 37

drvinceknight · 2017-12-04T15:49:02Z

axelrod/strategies/axelrod_second.py

+
+        if turn == 38 and opponent.history[-1] == D and opponent.cooperations == 36:
+            self.mode = "Fair-weather"
+            # These flags would already be set from turn == 37 logic below.


I think we can remove these two lines of inline comments These flags ...lower this turn.

drvinceknight · 2017-12-04T15:57:08Z

axelrod/strategies/axelrod_second.py

+    Strategy submitted to Axelrod's second tournament by Paul Harringtn (K75R),
+    and came in eighth in that tournament.
+
+    This strategy has three modes:  Normal, and Fair-weather, Defect.


Could we also add a sentence along the lines of what you wrote in the pdf: that the Fair-weather mode occurs only in 1 very specific type of situation.

drvinceknight · 2017-12-04T15:59:29Z

axelrod/tests/strategies/test_axelrod_second.py

+        # The defect streak will always be detected from here on, because it
+        # doesn't reset.  This logic comes before parity streaks or the turn-
+        # based logic.
+        self.versus_test(axelrod.Defector(), expected_actions=actions)


Another thought, could we add a few attrs={"mode": "Normal", "recorded_defects": ...} to the versus_test call so that we specifically test some of these attribute values? Not necessarily overkill but at least the mode and perhaps one or two others as you see fit?

That's cool. Didn't know that was possible.

gaffney2010 · 2017-12-05T07:42:00Z

This new error is confusing to me. Any idea?

drvinceknight · 2017-12-05T09:06:37Z

axelrod/strategies/axelrod_second.py

+            else:
+                self.exit_defect_meter -= 3
+            if self.exit_defect_meter >= 11:
+                self.mode = "Normal"


Coverage is failing because we don't have a test case that runs this code block:

axelrod/strategies/axelrod_second.py 436 4 99% 1196-1199

(That's the output of coveralls showing that lines 1196-1199 are not hit.)

Oh, interesting! I'll work on it.

marcharper

Thanks for this contribution! Since the code is complicated some more inline comments would be helpful.

@drvinceknight @meatballs I did not try to verify the logic in comparison to the Fortran source.

marcharper · 2017-12-08T01:37:34Z

axelrod/strategies/axelrod_second.py

+    In Normal and Fair-weather modes, the strategy begins by:
+
+    - Update history
+    - Detects random if turn is multiple of 15 and >=30.


Detects random opponent? (Not overly fussed)

marcharper · 2017-12-08T01:39:51Z

axelrod/strategies/axelrod_second.py

+    parity streak that we're pointing to.  If the parity streak that we're
+    pointing to is then greater than `parity_limit` then reset the streak and
+    cooperate immediately.  `parity_limit` is initially set to five, but after
+    its been hit eight times, it decreases to three.  The parity streak that


marcharper · 2017-12-08T01:40:16Z

axelrod/strategies/axelrod_second.py

+    pointing to is then greater than `parity_limit` then reset the streak and
+    cooperate immediately.  `parity_limit` is initially set to five, but after
+    its been hit eight times, it decreases to three.  The parity streak that
+    we're pointing to also gets incremented if in normal mode and WE defect but


marcharper · 2017-12-08T01:40:35Z

axelrod/strategies/axelrod_second.py

+    cooperate immediately.  `parity_limit` is initially set to five, but after
+    its been hit eight times, it decreases to three.  The parity streak that
+    we're pointing to also gets incremented if in normal mode and WE defect but
+    not on turn 38, unless the result of a defect streak.  Note that the parity


unless we are defecting as a result of defect streak?

Right. I'll clarify.

marcharper · 2017-12-08T01:40:45Z

axelrod/strategies/axelrod_second.py

+    its been hit eight times, it decreases to three.  The parity streak that
+    we're pointing to also gets incremented if in normal mode and WE defect but
+    not on turn 38, unless the result of a defect streak.  Note that the parity
+    streaks reset but the defect streak doesn't.


marcharper · 2017-12-08T01:47:28Z

axelrod/strategies/axelrod_second.py

+        return False
+
+    def detect_parity_streak(self, last_move):
+        self.parity_bit = 1 - self.parity_bit # Flip bit


docstring please

marcharper · 2017-12-08T01:47:42Z

axelrod/strategies/axelrod_second.py

+        return False
+
+    def detect_parity_streak(self, last_move):
+        self.parity_bit = 1 - self.parity_bit # Flip bit


two spaces before # (and several more below)

marcharper · 2017-12-08T01:48:57Z

axelrod/strategies/axelrod_second.py

+        if self.history[-1] == D:
+            self.history_row += 2
+
+        # If generous 2 turn ago and opponent defected last turn


2 turns ago

marcharper · 2017-12-08T01:49:41Z

axelrod/strategies/axelrod_second.py

+            self.history_row += 2
+
+        # If generous 2 turn ago and opponent defected last turn
+        if self.generous_n_turns_ago == 2 and opponent.history[-1] == D:


Does "generous_n_turns_ago" mean "last generous n turns ago" or something else? It's ambiguous IMO.

Yes. I can add "last_" here.

marcharper · 2017-12-08T01:50:02Z

axelrod/strategies/axelrod_second.py

+        if self.mode == "Fair-weather":
+            if opponent.history[-1] == D:
+                self.mode = "Normal" # Post-Defect is not possible
+                #Continue below


Remove this comment

I'm trying to express here that this is the only place in the code where, following a mode-switch, we don't immediately return a value. Instead we actually treat this turn as a "Normal" mode code. I'll take your advice if you think it's really not worth mentioning. (Or if I should say more clearly.)

I don't feel strongly about it but it's clear that the code continues to run IMO

gaffney2010 · 2017-12-08T02:04:35Z

@marcharper I'll make these changes and look for places where an in-line could be helpful. Thanks.

marcharper · 2017-12-08T02:08:57Z

Thanks! Some of the Fortran strategies are hard to understand so the extra inline comments are really helpful and will future-proof the code more completely. We don't want to repeat the current situation where the original strategies have very little documentation and we can't really know if they are implemented correctly. For example, there's a very rarely triggered behavior in this strategy which suggests that something wasn't implemented as intended (but we can't really know).

drvinceknight · 2017-12-08T08:04:22Z

@drvinceknight @meatballs I did not try to verify the logic in comparison to the Fortran source.

I believe this is the one last thing that needs to be done review wise apart from the requested changes from @marcharper 👍

gaffney2010 · 2017-12-08T08:09:02Z

I did not try to verify the logic in comparison to the Fortran source.

I think (hope) my attached PDF in original PR can be helpful here.

drvinceknight · 2017-12-08T08:13:05Z

I think (hope) my attached PDF in original PR can be helpful here.

yup it really did, to be clear I've gone through it myself and am happy, coupled with the agreement in the fingerprints I'm certain it's implemented correctly. Just in the case of this strategy which is pretty complex (kudos to what must have been a big translation effort @gaffney2010!) I think having a second set of eyes confirm would be good :)

gaffney2010 · 2017-12-08T08:28:44Z

In this first commit just now, I modified the comments per Marc. In the second commit ("Cleaned up record history logic"), I very slightly modified the record history logic. [Because in the last version I was updating move_history after the player enters and leaves Defect mode, the same way that the Fortran strategy does (which gets complicated after leaving Defect mode). But I realized that since you can only enter Defect mode once, this matrix isn't being used again, so I just made it stop recording at that point, and was able to clean stuff up.] Because it was a logic change. I re-ran all the tests/FPs. It still looks good, and I can upload the FPs if needed.

drvinceknight · 2017-12-08T08:32:51Z

It still looks good, and I can upload the FPs if needed.

Sounds good, if you could re upload just so we have record here that'd be great 👍

gaffney2010 · 2017-12-08T08:34:32Z

FPs. [The second in each pair is the Python run ones. Only these were updated.]

drvinceknight · 2017-12-08T15:13:53Z

axelrod/strategies/axelrod_second.py

+                    chi_squared += (expect - self.move_history[i, j]) ** 2 / expect
+
+        # Caching value only for testing purposes, not used otherwise
+        self.chi_squared = round(chi_squared, 3)


So it doesn't look like you've gone with the suggestion of making the chi_squared calculation it's own function which could be tested?

You could essentially keep this exactly as it is:

+ def calculate_chi_squared(self): + expected_matrix = np.outer(self.move_history.sum(axis=1), + self.move_history.sum(axis=0)) / denom + + chi_squared = 0.0 + for i in range(4): + for j in range(2): + expect = expected_matrix[i, j] + if expect > 1.0: + chi_squared += (expect - self.move_history[i, j]) ** 2 / expect + + # Caching value only for testing purposes, not used otherwise + return chi_squared

Then in detect_random you just use self.calculate_chi_squared but then, for the test, you can directly run a match for your strategy and call the self.calculate_chi_squared method to ensure it gives the correct value (with no need for the self.chi_squared.

My other suggestion was to have this as a completely static method/function but it's also fine to leave it as a method of the class.

I didn't make the change, because it sounded like it was optional AND because I'm still not seeing how, under this alternative approach, I would test that move_history is getting filled out right. If I understand your comment, it was that, this doesn't matter because I can test the Chi-Squared logic on my matrix by manually entering the matrix:

self.assertEqual(round(axelrod_second.calculate_chi_squared(np.matrix([[0, 2], [5, 6], [3, 6], [4, 2]])), 3), 2.395)

While I agree that this would successfully test the chi-squared logic. How do I test that move_history correctly incremented to [[0, 2], [5, 6], [3, 6], [4, 2]] ?

I'm not sure I see the connection between testing the move_history and calculate_chi_squared. As far as I can tell you're not testing that in the tests anyway:

+ actions += [(D, C)] + self.versus_test(axelrod.Random(0.5), expected_actions=actions, seed=10, attrs={"chi_squared": 2.395}) + # The history matrix will be [[0, 2], [5, 6], [3, 6], [4, 2]]

You can keep this as a sub method of the class if you prefer in which case nothing here changes except you no longer need to artificially set self.chi_squared as an attribute.

So that test (which if I understand correctly is the only purpose of this rounded chi_squared attribute which we'd rather not need) would be modified to be an actual match. Something like (I haven't run this):

axl.seed(10) player = self.player() match = axl.Match((player, axl.Random()), turns=len(expected_actions)) actions = match.play() self.assertEqual(actions, expected_actions) # Just to be consistant with the current test. self.assertEqual(round(player.calculate_chi_squared()), 2.395)

I'm not sure I see the connection between testing the move_history and calculate_chi_squared.

calculate_chi_squared uses move_history. You can get 2.935 precisely if move_history is filled out (correctly) as [[0, 2], [5, 6], [3, 6], [4, 2]]

I think I understand better now...

self.assertEqual(round(player.calculate_chi_squared()), 2.395)

I wasn't making the connection that the player would still be in the end-of-turn-30 state after you ran the axl.Match(), and I was trying to figure out how I could check everything in the Match() function, because I was imagining that the player would disappear afterwards. [I guess the way I wrote it, it would disappear.]

I'll change it like you have it. Thanks.

calculate_chi_squared uses move_history. You can get 2.935 precisely if move_history is filled out (correctly) as [[0, 2], [5, 6], [3, 6], [4, 2]]

Cool, so we can either:

create compute_chi_squared as it's own function (outside the class) that in the test gets passed np.array([[0, 2], [5, 6], [3, 6], [4, 2]]), turn as well as recorded_defects (which I believe it also uses). So compute_chi_squared would be a function that takes two arguments which is in turn called by detect_random. Something like:

def detect_random(self, turn): chi_squared = compute_chi_squared(self.move_history, self.recorded_defects) if chi_squared > 3 ...

The test would be:

self.assertEqual(round(axelrod_second.calculate_chi_squared(np.matrix([[0, 2], [5, 6], [3, 6], [4, 2]])), ?, ?), 2.395) # Replace ? with the correct turn recorded_defects :)

Create a method which means we then need to test using a match.

I think option 1 is probably the cleanest way of doing this but 2 would do the trick also :)

I think I understand better now...

Cool 👍

marcharper · 2017-12-09T23:11:09Z

axelrod/tests/strategies/test_axelrod_second.py

+        actions += [(D, D)] * 14
+        # Mutual defect for a while, then exit Defect mode with two coops
+        actions += [(C, D)] * 2
+        self.versus_test(Rand_Then_Def, expected_actions=actions, seed=10, \


Don't need the slash here

marcharper · 2017-12-09T23:12:37Z

axelrod/tests/strategies/test_axelrod_second.py

+        'manipulates_state': False
+    }
+
+    def test_strategy(self):


Thanks for moving the chi-squared attribute. This test function is pretty long now, you might consider breaking it up into several test functions. (optional)

You could also target specific functions with their own tests, e.g. for detect_streak and friends.

marcharper · 2017-12-09T23:12:57Z

axelrod/tests/strategies/test_axelrod_second.py

+        actions = [(C, D), (D, C)] + [(C, C)] * 34 + [(D, C)]
+        # Two cooperations scheduled after the 37-turn defection
+        actions += [(C, C)] * 2
+        # TFT twice, then low-random # yields a DCC combo.


Did you want two lines for this comment?

marcharper · 2017-12-09T23:13:09Z

axelrod/tests/strategies/test_axelrod_second.py

+        # TFT twice, then low-random # yields a DCC combo.
+        actions += [(C, C)] * 2
+        actions += [(D, C), (C, C), (C, C)]
+        # Don't draw next random # until now.  Again DCC.


marcharper · 2017-12-09T23:13:44Z

axelrod/tests/strategies/test_axelrod_second.py

+        actions += [(D, C), (C, D), (D, C), (C, D), (C, C)]
+        # This is the seventh time we've hit the limit.  So do it once more.
+        actions += [(C, D), (D, C), (C, D), (D, C), (C, D), (C, C)]
+        # Now hit thi limit sooner


thi --> this?

marcharper · 2017-12-09T23:16:48Z

axelrod/tests/strategies/test_axelrod_second.py

+        # The history matrix will be [[0, 2], [5, 6], [3, 6], [4, 2]]
+        actions = match.play()
+        self.assertEqual(actions, expected_actions)  # Just to be consistant with the current test.
+        self.assertEqual(round(player.calculate_chi_squared(len(expected_actions)),3), 2.395)


space after comma (before 3); drop the round or use assertAlmostEqual (which allows you to specify a number of places for equality):

self.assertAlmostEqual(a, b, places=4)

Link to the docs for assertAlmostEqual.

marcharper

Approving; I've asked for some minor additional changes.

drvinceknight · 2017-12-10T11:17:09Z

axelrod/tests/strategies/test_axelrod_second.py

+        match = axelrod.Match((player, axelrod.Random()), turns=len(expected_actions))
+        # The history matrix will be [[0, 2], [5, 6], [3, 6], [4, 2]]
+        actions = match.play()
+        self.assertEqual(actions, expected_actions)  # Just to be consistant with the current test.


Can we remove # Just to be consistant with current test (that was just meant as a note to you in my comment but here it could be confusing).

drvinceknight · 2017-12-10T11:18:07Z

Thanks for doing everything, this is looking great now, let's just remove that one inline comment which could be confusing later down the line :) Then this is good to go! 👍

drvinceknight · 2017-12-10T12:51:14Z

Awesome: thank you for making all the requested tweaks. 👍

gaffney2010 added 2 commits December 4, 2017 09:19

Implemented Harrington (k75r from Axelrod's Second)

6b24094

Merge branch 'master' of https://github.com/Axelrod-Python/Axelrod

d902e41

drvinceknight requested changes Dec 4, 2017

View reviewed changes

gaffney2010 added 3 commits December 5, 2017 01:39

Minor formatting/comment fixes.

5e4f4c8

Added some attrs arguments to the test file.

ffd44a8

Fixed indentation in docstrings.

8231f07

drvinceknight reviewed Dec 5, 2017

View reviewed changes

Added tests for Defect mode.

a8db7da

marcharper requested changes Dec 8, 2017

View reviewed changes

gaffney2010 added 2 commits December 8, 2017 03:50

Edited and added comments/docstrings.

5bf2d69

Cleaned up record history logic.

72f5edb

meatballs approved these changes Dec 8, 2017

View reviewed changes

drvinceknight requested changes Dec 8, 2017

View reviewed changes

Moved chi-squared to its own function.

15cb4a2

marcharper reviewed Dec 9, 2017

View reviewed changes

marcharper approved these changes Dec 9, 2017

View reviewed changes

gaffney2010 added 2 commits December 9, 2017 18:22

Minor changes per PR request feedback.

b9c0b35

Fixed places on assertAlmostEqual

a81b750

drvinceknight reviewed Dec 10, 2017

View reviewed changes

Update test_axelrod_second.py

bf2c855

drvinceknight approved these changes Dec 10, 2017

View reviewed changes

drvinceknight merged commit 3402156 into Axelrod-Python:master Dec 10, 2017

drvinceknight mentioned this pull request Dec 13, 2017

After next (>v3.10.0) release: strategies to update Axelrod-Python/axelrod-fortran#72

Closed

gaffney2010 mentioned this pull request Jan 17, 2018

K75R Strategy #1101

Closed

Implemented Harrington (k75r from Axelrod's Second) #1146

Implemented Harrington (k75r from Axelrod's Second) #1146

Conversation

gaffney2010 commented Dec 4, 2017

drvinceknight left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gaffney2010 commented Dec 5, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

marcharper left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gaffney2010 commented Dec 8, 2017

marcharper commented Dec 8, 2017

drvinceknight commented Dec 8, 2017

gaffney2010 commented Dec 8, 2017

drvinceknight commented Dec 8, 2017 • edited Loading

gaffney2010 commented Dec 8, 2017

drvinceknight commented Dec 8, 2017

gaffney2010 commented Dec 8, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

drvinceknight Dec 8, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

drvinceknight Dec 8, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

marcharper left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

drvinceknight commented Dec 10, 2017

drvinceknight commented Dec 10, 2017

drvinceknight commented Dec 8, 2017 •

edited

Loading

gaffney2010 commented Dec 8, 2017 •

edited

Loading

drvinceknight Dec 8, 2017 •

edited

Loading

drvinceknight Dec 8, 2017 •

edited

Loading