Implemented Black, K83R from Axelrod's Second

[gaffney2010]

Fingerprints Below

[drvinceknight]

All looks fantastic. Thanks!

I have looked through https://github.com/Axelrod-Python/TourExec/blob/master/src/strategies/k83r.f and confirm the logic is implemented correctly. 👍

[marcharper]

Are we sure the memory depth is 5? It's a weird case since you need to know the round number + the last five rounds.

[drvinceknight]

Are we sure the memory depth is 5? It's a weird case since you need to know the round number + the last five rounds.

This is similar to memory one strategies that all have a memory depth of 1 but that act differently on their first move to any other right?

[gaffney2010]

You don't need to know the full round number. If you gave me just the last five moves as a list, I could check if that list is length 5 or not.

[drvinceknight]

You don't need to know the full round number. If you gave me just the last five moves as a list, I could check if that list is length 5 or not.

Even better point 👍

[marcharper]

I guess that's what I'm asking -- on the first move there's no history so it's sort of a special case and players must have a separate rule. Here on rounds say 4, 5, 6 you have to know the round number and then decide from one of two behaviors. So we need to know if there are at least 5 rounds of history already, so we need the size of the history and therefore all of the history. Or is the rule that we can ask "Are there at least X rounds of history?" and then take the minimum needed value of X as the memory depth?

[marcharper]

Also think about the case where there's X rounds of initial play and then a behavior that depends on the last Y rounds, where X != Y.

[marcharper]

Also, if the implementation were setting a bit if_round_greater_than_five then we usually call that kind of situation infinite memory...

[drvinceknight]

Here on rounds say 4, 5, 6 you have to know the round number and then decide from one of two behaviors. So we need to know if there are at least 5 rounds of history already, so we need the size of the history and therefore all of the history. Or is the rule that we can ask "Are there at least X rounds of history?"

It's a good question. I think that my personal preference would be to define memory as "if you give history[-memory:] or history you get the same behaviour". I believe this isn't the first strategy that poses this question so I'd suggest we merge as is (based on the mem 1 behaviour/convention it feels right to me that this has mem depth 5). And perhaps aim to implement a test for memory on all players?

[gaffney2010]

In this sense, Tit-For-Tat has to know the round number, since there's a special rule for when turn=1, so you have to know the turn number. But I think it's clear that TFT is memory=1; by analogy, this is memory=5. And I think it's intuitive; you're allowed to be forgetful.

[drvinceknight]

"if you give history[-memory:] or history you get the same behaviour"

In this particular case if we're on the third turn with opponent.history=[C, C, C] then opponent.history[-5:] would also be [C, C, C], so:

+        if len(opponent.history) < 5:
+            return C

would return C.

And perhaps aim to implement a test for memory on all players?

This would potentially assist with #597 (not in finding an algorithm but at least verifying upper bounds on memory lengths).

[drvinceknight]

In this particular case if we're on the third turn with opponent.history=[C, C, C] then opponent.history[-5:] would also be [C, C, C], so:

If we're on the 150th turn then len(opponent.history[-5:]) is 5 so the rest of the logic kicks in.

[drvinceknight]

I've opened #1156 as a suggestion for a test (to perhaps move the discussion about the way to go over there).

[marcharper]

Ok, let's discuss elsewhere. I'll merge as is.