3 strategies that are implemented in axelrod do not agree

[drvinceknight]

In #30 I have written a test that checks implemented deterministic strategies.

3 are not agreeing:

• https://github.com/Axelrod-Python/TourExec/blob/v0.2.0/src/strategies/k57r.f - Axelrod: Nydegger
• https://github.com/Axelrod-Python/TourExec/blob/v0.2.0/src/strategies/K59R.f - Axelrod:
  RevisedDowning
• https://github.com/Axelrod-Python/TourExec/blob/v0.2.0/src/strategies/k86r.f - Axelrod: Grofman

Grofman is perhaps a simple case to investigate: it is described and implemented in the Axelrod library as a stochastic strategy but it's classified as deterministic in axelrod_fortan...

[meatballs]

Well, isn't this getting interesting?!!

[drvinceknight]

Indeed :)

[meatballs]

Grofman is indeed stochastic in the Axelrod library. It returns a random choice in its last line.

But, having looked through its code, K86R is defintely deterministic.

[drvinceknight]

It's implemented based on the definition in Axelrod's first paper (which is now in the docstring). There's a chance that the strategy Grofman wrote for the second was intended to be different perhaps.

[meatballs]

They start in a similar fashion, but these are not the same strategy

[drvinceknight]

In which case I'd suggest that we modify http://axelrod.readthedocs.io/en/stable/reference/overview_of_strategies.html#id15 to indicate that it's implemented in axelrod? But let's make sure hey? I've just read through the paper again and I don't see anything that indicates that it is intended to be the same strategy, just the same author.

[meatballs]

I assume there is a missing 'not' in there, in which case, I agree.

I think he replaced the random choice used in the first tournament with some deterministic code for the second

[drvinceknight]

I assume there is a missing 'not' in there, in which case, I agree.

Yup. "not implemented"

I think he replaced the random choice used in the first tournament with some deterministic code for the second

Could be the case with the other two also: they modified their strategies...

[meatballs]

Comment in line 3 of K59R:

C TYPED BY AX, 3/27/79 (SAME AS ROUND ONE REV.DOWNING)

[meatballs]

and in K57R:

C TYPED BY AX, 3/27/79 (SAME AS ROUND ONE NYDEGR)

[meatballs]

K57R versus Nydegger: Just going by the comments in the code, I think the values it uses for defection are different - same logic but using different values.

However, following the archaic conditionals through will take some effort to be certain.

[marcharper]

Apparently there are two Grofman's (Mark and Bernard). Are the known differences all deterministic? If not it could be simply a difference in randomizations.

Otherwise I'm not surprised that a few are different -- there were definitely cases where we had to make choices knowing that the text descriptions were incomplete.

[drvinceknight]

Apparently there are two Grofman's (Mark and Bernard). Are the known differences all deterministic? If not it could be simply a difference in randomizations.

• Grofman's strategy in Axelrod's first tournament was stochastic. It was not in his second.

Otherwise I'm not surprised that a few are different -- there were definitely cases where we had to make choices knowing that the text descriptions were incomplete.

I think this is just highlighting that we've made a bit of a jump/assumption with http://axelrod.readthedocs.io/en/stable/reference/overview_of_strategies.html#axelrod-s-second-tournament

The strategies from Axelrod's first tournament are all implemented based on the text description provided in https://www.jstor.org/stable/173932

We then assumed that if a strategy was submitted to the second tournament by the same author that it was the same strategy. That is not made expicit in Axelrod's second paper https://www.jstor.org/stable/173638. There is sadly no text description of the strategies in that second paper so the Fortran code is the description of the strategy.

So I think it's more of an exception than a rule when the strategy is the same. I'm going to open PRs both here and for the tournament docs to mark most of these as "not" implemented (the ones that pass the test and Random I think are safe to assume are the same). If and when there are implemented (or we know that they are implemented) we should change that.

I'd also suggest that when we implement them we don't call them by their author name but by the name of the Fortran file (including author name in the title etc), or we call them MoreGroffman (or something like that... More because that's the extra word in the paper about Axelrod's second tournament).

[drvinceknight]

On Axelrod: Axelrod-Python/Axelrod#1090
Here: #33

[drvinceknight]

Some quotes from Axelrod's second paper which I think back up that there's no reason to assume that the same author submitted the same strategy:

The game theorists who participated in the first round of the tournament were also invited to try again.

So not automatic.

The entrants also drew their own lessons from the experience of the first round.

So strategies did change?

[drvinceknight]

I think we can close this.