There is a fair possibility that a number of programs whose dominant strategy is “cooperate with other dominant cooperators, do something else otherwise” (what ThrustVectoring calls AbsolutismBot “wrapper”) will end up sharing the top spot with the same number of points. I wonder if it makes sense to discourage this approach in some way.
That depends on whether players get utility from PD payoffs or from winning the tournament. A clique that wants a high total payoff will use mutual cooperation. A clique that wants to grab the #1 spot will have all players sacrifice to one designated player.
I’m pretty sure that incorporating code written by someone else into your entry qualifies. I think the highest single entry might be to cheat and make everyone think you are in that tribe but defect anyway, or it might be dominant to defect against any program displaying tribal affiliations (other than this new tribe, of course). The dominant tribe is the tribe with the most members and best tribal identification, not the tribe with the best way of judging an opponents intentions.
There is a difference between a “tribe system” as mentioned by yourself and one person winning by submitting 1000 entries. The goal as I understand it is simply to maximize your score by whatever means possible, not accurately guess your opponents intentions.
It is certainly fair, but in a situation where, say, there is only one escape pod available, you still have to fight it out with others. What I had in mind is to make this more adversarial if the obvious and boring approach is full or nearly full cooperation.
Sorry, I do not follow. Are you saying that there is a guaranteed best strategy in the winner-takes-all case? And that “cooperate with clones, defect against everyone else” is it?
“Winner takes all” doesn’t have “do better”. You either take it all or you have nothing at all. If you didn’t win you didn’t do any better than a painted rock.
I think we don’t mean the same thing when we say “winner takes all”. (BTW, I only just noticed you used that phrase in reply to a comment where Eliezer uses “zero-sum games”, which is not at all the same thing. In a ZSG, me and a buddy can band together, steal your money, and split between ourselves. In WTA I have to fight my ally, I can’t share with him.)
I probably misunderstand what you mean. People temporarily cooperate all the time even if all of them know that in the end “there can be only one”. It may be advantageous to band with your near-clones and hope that you are better than them when dealing with others, thus gaining a decisive edge.
OK, I got it. (It helps if you look at the single thread, and click on “Show more comments above.” to see the entire chain.)
This contest is supposed to rank strategies for iterated PD. Your one-escape-pod comment suggests you are interested in a contest that finds the winner stategy when there can be a single winner. That would be a very different kind of contest, though supperficially similar.
Obviously, a ranking contest can be transformed into a winner-picking contest, by declaring the first-ranked candidate the winner, if there is a mechanism for enforcing the lack of ties for first place. This mechanism is not necessarily easy to add unless the contest is designed for it from the beginning, and since there are lots of possible ranking methods it is probably often impossible.
But due to the superficial similarity (and perhaps an abuse of technical language), I think the distinction between what this contest does and what you wanted it to do was lost somewhere in your initial exchange with Eliezer.
Of course, what you say about temporary cooperation is perfectly correct. I think our terminology misunderstanding stems from the perspective, I was more focused on the entire contest (which is more adversarial, as the goal is the be ranked above others), and I believe you were focusing on the individual encounters (which are slightly more cooperative, since mutual destruction is very bad). We probably don’t disagree on what the words mean, we were just thinking of different issues.
(This is a constant danger when discussing games composed of other games (e.g. the iterated tournament), so we should probably have been more explicit. In retrospective Eliezer using the word “fair” should have been a warning everyone is talking about different things...)
There is a fair possibility that a number of programs whose dominant strategy is “cooperate with other dominant cooperators, do something else otherwise” (what ThrustVectoring calls AbsolutismBot “wrapper”) will end up sharing the top spot with the same number of points. I wonder if it makes sense to discourage this approach in some way.
That depends on whether players get utility from PD payoffs or from winning the tournament. A clique that wants a high total payoff will use mutual cooperation. A clique that wants to grab the #1 spot will have all players sacrifice to one designated player.
Why isn’t this just a fair result?
I’m pretty sure that incorporating code written by someone else into your entry qualifies. I think the highest single entry might be to cheat and make everyone think you are in that tribe but defect anyway, or it might be dominant to defect against any program displaying tribal affiliations (other than this new tribe, of course). The dominant tribe is the tribe with the most members and best tribal identification, not the tribe with the best way of judging an opponents intentions.
There is a difference between a “tribe system” as mentioned by yourself and one person winning by submitting 1000 entries. The goal as I understand it is simply to maximize your score by whatever means possible, not accurately guess your opponents intentions.
It is certainly fair, but in a situation where, say, there is only one escape pod available, you still have to fight it out with others. What I had in mind is to make this more adversarial if the obvious and boring approach is full or nearly full cooperation.
I don’t think anyone disputes that in zero-sum games you play the Nash equilibrium move.
Sorry, I do not follow. Are you saying that there is a guaranteed best strategy in the winner-takes-all case? And that “cooperate with clones, defect against everyone else” is it?
In winner takes all, you can’t cooperate and win, because “cooperate” implies you didn’t take it all.
You can still do better than your almost-clones if the difference is in how you play detected non-clones or non-clones misdetected as almost-clones.
“Winner takes all” doesn’t have “do better”. You either take it all or you have nothing at all. If you didn’t win you didn’t do any better than a painted rock.
I think we don’t mean the same thing when we say “winner takes all”. (BTW, I only just noticed you used that phrase in reply to a comment where Eliezer uses “zero-sum games”, which is not at all the same thing. In a ZSG, me and a buddy can band together, steal your money, and split between ourselves. In WTA I have to fight my ally, I can’t share with him.)
I probably misunderstand what you mean. People temporarily cooperate all the time even if all of them know that in the end “there can be only one”. It may be advantageous to band with your near-clones and hope that you are better than them when dealing with others, thus gaining a decisive edge.
OK, I got it. (It helps if you look at the single thread, and click on “Show more comments above.” to see the entire chain.)
This contest is supposed to rank strategies for iterated PD. Your one-escape-pod comment suggests you are interested in a contest that finds the winner stategy when there can be a single winner. That would be a very different kind of contest, though supperficially similar.
Obviously, a ranking contest can be transformed into a winner-picking contest, by declaring the first-ranked candidate the winner, if there is a mechanism for enforcing the lack of ties for first place. This mechanism is not necessarily easy to add unless the contest is designed for it from the beginning, and since there are lots of possible ranking methods it is probably often impossible.
But due to the superficial similarity (and perhaps an abuse of technical language), I think the distinction between what this contest does and what you wanted it to do was lost somewhere in your initial exchange with Eliezer.
Of course, what you say about temporary cooperation is perfectly correct. I think our terminology misunderstanding stems from the perspective, I was more focused on the entire contest (which is more adversarial, as the goal is the be ranked above others), and I believe you were focusing on the individual encounters (which are slightly more cooperative, since mutual destruction is very bad). We probably don’t disagree on what the words mean, we were just thinking of different issues.
(This is a constant danger when discussing games composed of other games (e.g. the iterated tournament), so we should probably have been more explicit. In retrospective Eliezer using the word “fair” should have been a warning everyone is talking about different things...)