Wow, I made it to the top three!

In the sake of fairness, however, I have to say that I think you made a mistake with my submission: I had first submitted SimpleTFTseer and then asked you to replace it with VeryOpportunisticFarseerBot, but it seems that you have entered both, so I had two bots in the tournament.
SimpleTFTseer did poorly and wasn't designed to advantage VeryOpportunisticFarseerBot, hence this probably didn't make a significant difference, but I think it is right to disclose it.

Anyway, thanks for your effort, it has been fun to participate!

My primary design criterion was to get the maximum number of points possible vs as many opponents as possible, so I'm particularly pleased that CheeseBot gets the highest average score in the first round by quite a large margin.

I wanted to make by bot "unexploitable" in the sense that no bot could ever bring me below the C/C equilibrium of 300 points without hurting its own score in the process. In this regard my bot still wasn't perfect, but as far as I can see no other bot was able to exploit me in this sense.

If the goal was to maximise IPD performance I would have won outright, but as was noted in the original thread the game we are playing is not a two-player non-zero sum IPD, but rather a much more complicated zero-sum game with many players.

For example, TatForTit defected on the 3rd last turn when playing against me, causing me to lose 5 points at the cost of 2 points. In the IPD this would be a stupid move, but in the tournament setting this is a pretty good deal.

However, this still doesn't explain everything; if my bot worked correctly, it would've defected on the 4th last turn if it anticipated being defected against on the 3rd last turn. I'm not sure why this failed, so I'll have to look into the CheeseBot vs TatForTitBot matchup in greater detail to work out what was going on there.

Any reason I shouldn't move to Main and Promote?

Joint first! \o/

Clarification for DMRB: a "dumb" bot is one that doesn't try to simulate vis opponent. If DMRB plays against a not-dumb opponent, ve plays tit-for-tat. Against a dumb opponent, ve simulates the future game tree a few levels deep to figure out how to get the most points. (To distinguish dumb from not-dumb: assume they're dumb, and run the simulation under a time limit, pitting them against TimeoutBot.)

There were only three dumb bots in this tournament (AnderBot, DefectBot, SimHatingTitForTat), so it seems like TitForTatBot would have done approximately as well. I expected the exploitation code to be mostly useful in early rounds, I hadn't realised that the sample bots wouldn't be taking part.

I considered defecting in the last round, but DMRB was (by design) sufficiently transparent that an opponent could have anticipated me doing that. It looks like if I'd done that, SimpleTFTSeerBot and DMRB would both have lost 300 points in the first round, and DMRB would have been knocked out. (Unless STSB already considered DMRB to time out.) Interestingly, I did defect in the last round against dumb bots, which would have been even easier to anticipate, but it doesn't look like anyone punished me for that.

(edit: actually only 200 points, not 300, but that would still have knocked me out.)

(edit 2: looking at the results, STSB does defect against DMRB, so defecting in the final round wouldn't have cost me anything there. In the final round, STSB is dumb. In the first round, STSB simulates vis opponent playing verself in the final round, and defects unless vis opponent cooperates. Since DMRB defects in that situation, STSB defects in the first round, and we fall into a pit of (D,D). So final-round defection against dumb bots cost me versus STSB, and final-round not-automatic-defection against clever bots didn't gain me anything against ver.)

Since people seem to be unsure about this, I figured it would be good to make it clear that there are many simple strategies that are provably symmetric Nash Equilibria and cooperate 1v1 vs relatively broad varieties of opponents (unlike a CliqueBot in the non-iterated source-sharing version). Here's one example:

If either you or your opponent has ever defected in the past, Defect.
Otherwise, if it's the 98th round and you simulate that your opponent will defect on the 99th round if you both cooperate on the 98th round, or if you simulate that your opponent will defect on the 100th round if you both cooperate on the 98th and 99th rounds, or if your simulation attempt times out, Defect.
Otherwise, Cooperate.

When playing against this bot, if you defect on the 98th round or earlier, you will be defected against for the rest of the game, and mutual defection loses you at least 4 points which outweighs your 2-point gain for DC. If you cooperate for 98 rounds and then defect on the 99th, 100th, or both, you will be defected against for the last three rounds and hence you will lose points.

Consequently, the only way to maximise your reward vs this bot is to Cooperate on every round, which gets you 300 points, so this bot is unexploitable in the same way that PrudentBot in the non-iterated source-sharing PD. Also, this bot clearly cooperates for all 100 rounds if it plays against itself.

The basic starting point for this type of bot is to punish defections in past rounds via simple Tit-for-Tat-like methods. This only leaves one basic flaw - Tit-for-Tat can be exploited via defection on the last round. The solution to this problem is to patch this hole up by simulating future rounds (with the same matchup) and punishing them in advance. Additionally, in any matchup between two bots that use this general principle, the recursive simulation process is guaranteed to eventually hit a base case.

I think there are quite a large number of distinguishable cooperative symmetric equilibrium strategies, but I think they would all follow this general principle of punishing past defections as well as defections detected in simulated future rounds.

I didn't go for something quite this simple because the tournament setup means the point of the game is not simply to maximise score, and because I wanted to perform optimally vs some simple bots like DefectBot, CooperateBot and RandomBot.

Would a tournament where the bots can see how other bots behave with each other be possible?

The simplest version would be that bots just see all the interactions. It could be made more realistic if the bots have memory limits, and much more realistic (and much messier) if bots have some ability to conceal their interactions.

"Do not attempt long chains of reasoning or complicated plans."

I find it somewhat amusing that SimHatingTitForTat has an average score second only to CheeseBot, and is one of the constant first-round survivalists... but just doesn't quite make it into the finals. Not bad for an algorithm that was an attempt to translate Tit-for-Tat's simplicity and reflectiveness into this particular tournament's terms. If the terms of the tourney were somewhat different, focusing more on 'avoiding total loss' rather than 'maximized score each game', then, well... :)

(I would be quite happy to see any further tournaments get run - I might even manage to come up with further ideas for the same.)

If you ran this tournament again, I would happily participate again!

This is brilliant work and a great summary and analysis of the whole thing.

I'd love to see a version where the bots do not get each other's code to use as a simulation, and are limited to their own experiences, but -- before making the cooperate-or-defect decision -- can pay the opponent a price that the opponent sets for the opponent's full history to that point. This would require three decisions, instead of one: decide what price to set for your own history, decide whether to pay the price for your opponent's history, and decide whether to cooperate or defect.

As a further variation, I'd love to see the same thing where contestants are allowed to submit multiple bots, so they can use some as information gatherers (until they die off).

And I'd upvote the suggestion of a variable-length simulation where no one knows the number of rounds ahead of time.

Max L.

Also worth checking, out of interest: How does each of the bots perform when playing against itself?

Thank you very much for running this and your effort in writing up the results; I'm very happy with how CheeseBot did!

Now I just need to work out how exactly TatForTit managed to exploit me...

Is there a way to analyze this tournament's data, if the final turns' scores were to be ignored? (That is, if any 'defect on the last turn' instructions didn't do anything useful in the games themselves, only for simulation purposes.)

Do I see a hint of a network-like effect in the results?

Simulate my opponent and figure out if defections will be punished; if so, Cooperate, and otherwise, defect. This allows bots following this strategy to always cooperate with each other, consistently providing them with a large number of points in every round, ensuring that they outcompete backstabbing or other aggressive strategies.

If I understand this, then the more bots that have this "cooperate unless defection is unpunished" strategy, the better the payoff for having that strategy. So in an evolutionary system, once a few bots get this strategy, it will be selected for with increasing selection pressure in subsequent generations.

My understanding of these tournaments is on the hairy edge of not existing, but if I have understood this correctly then this seems like a strong result explaining how a strategy might win in an evolutionary arms-race.

So, if I am reading this correctly, it looks like I (TatForTit) almost always won when I made it past the early bracket, but failed to do well in the early game. I actually somewhat predicted this. Oh well. Thanks for running this!

Something that bothers me about this tournament: I feel like a competitive tournament doesn't actually reward the kind of strategy that is meant to do well in Prisoner's Dilemna. As a (highly oversimplified) example, consider three bots who have the scores:

A: 10 B: 9 C: 2

Here, A is 'winning.' Suppose B can make a move that costs A 3 points and costs itself 1 point, leading to:

A: 7 B: 8 C: 2

B's payoff function has dropped. However, from a 'winning the tournament' approach, B has gone from 2nd to 1st, and so this outcome is now better for B. This feels wrong.

I doubt this was a really big issue here, but just on general principles I feel like competition by comparing scores is incompatible with a desire to explore the Prisoner's Dilemma, since you're turning a non-zero-sum game into a zero-sum game.

Dang it, days later and I'm still insanely curious to see how the results would differ if the length of the matches wasn't known by the algorithms. Either by removing the concept of limiting matches, or by ending matches far earlier(not just one or two steps) than they were 'planned' to end.

I've been pondering downloading the code, changing and running it, but my shoulder angels start slapping me at the thought of me neglecting my current projects to (even briefly) learn Haskell.

Just occurred to me that there's a way around this, I can offer a shameless bribe to someone else that already knows what they're doing and has some spare time. :-D

With that in mind, $20USD via paypal to the first person who runs the modified tournament and posts the results as a reply to this comment. If you don't want the cash, I'll donate it to MIRI/a charity of your choice, etc.

These results definitely confirmed my theory that my stupid bot which I thought about for 10 minutes got thoroughly crushed by anyone else's bot who thought about it for more than 10 minutes. :)

Grats to the winners!

I would find it very interesting if the tournament had multiple rounds and the bots were able to adapt themselves based on previous performance and log files they generated at runtime. This way they could use information like 'most bots take longer to simulate than expected.' or 'there are fewer cannon-fodder bots than expected' and become better adapted in the next round. Such a setup would lessen the impact of the fact that some bots that are usually very good underperform here because of an unexpected population of competitors. This might be hard to implement and would probably scare away some participants, though.

I'd like to add one more suggestion for the next version of this tournament:

1. If you have N bots, don't run NxN matches; instead, engineer their exchanges to be less complete, and more random. This may encourage the emergence of reputation.

And more generally, I would like to see systematic explorations in post-IPD tournaments; tournaments in which bots are able to (in constrained conditions) mold the utility landscape for their own ends.

Hey, my bot (TwoFacedTitBot) did pretty well for a while there! :) No bad, considering that I threw it together after reading one Haskell tutorial and on the basis of basically no analysis besides coming up with a bunch of variations on a theme and running them against each other.

The idea was supposed to be that the simulated round against MirrorBot would act as a crude test of whether 2FTB was being simulated (cause if the opponent was naively simulating me, then they would time out against MirrorBot due to infinite regress), and then if I was being simulated I would play nice and cooperate, and otherwise I would play the slightly more dickish TitForTat-but-defect-on-last strategy. I suspect that it only did that well because it was so similar to TitForTat, but nonetheless I'm at least a little pleased with myself.

I'd be interested to see how the tournament would run with default bots included, though. Like I think many people, I had somehow gotten it into my head that they would be included. And maybe RandomBot would improve the variance a little bit? :P

In any case, massive props to tetronian for running this Awesome Thing :)

Your win frequency graph is wrong; it should say 963 for VOFB, not 968.

Anyways, I did some more detailed statistics with the power of
ls | xargs perl -sn0e 'while (/next round:.(\N*)/sg) {print "$1->"}; print "$ARGV\n"' | sort -V
and here's the results I came up with:

963 matches went like this:
ALL->[Cheese,DMRB,SimHatingTFT,Switch,TwoFacedTit,VOFB]->[Cheese,DMRB,VOFB]

32 matches went like this:
ALL->[Cheese,DMRB,SimHatingTFT,Switch,TatForTit,TwoFacedTit]->[Cheese,DMRB,TatForTit]->[DMRB,TatForTit]->[TatForTit]
(13, 19, 30, 31, 33, 40, 42, 54, 74, 119, 137, 203, 236, 272, 301, 303,
321, 326, 343, 400, 438, 476, 516, 526, 539, 608, 626, 815, 823, 827, 941, 977)

3 matches went like this:
ALL->[Cheese,DMRB,SimHatingTFT,Switch,TatForTit,TwoFacedTit,VOFB]->[Cheese,DMRB,SimHatingTFT,TwoFacedTit]->[Cheese,DMRB]
(64, 553, 979)

Match #11 went like this:
ALL->[Cheese,DMRB,SimHatingTFT,Switch,TatForTit,VOFB]->[Cheese,DMRB,SimHatingTFT]->[Cheese,DMRB]

Match #12 went like this:
ALL->[Cheese,DMRB,SimHatingTFT,Switch,TwoFacedTit,VOFB]->[Cheese,DMRB,SimHatingTFT]->[Cheese,DMRB]

Would it be worthwhile to setup an automated, decentralized competition for iterated prisoner's dilemma bots, sort of like the roborumble for Robocode?

A request: Could somebody collate the data, and produce a table of the (average) scores for each program pairing?

"Simulate my opponent and figure out if it will defect or if defections will be punished; if so, Cooperate, and otherwise, defect."

Isn't that backwards?