Unnamed comments on Prisoner's Dilemma Tournament Results - Less Wrong

101 06 September 2011 12:46AM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Sort By: Best

Comment author: 06 September 2011 07:05:26PM 1 point [-]

I like it. One minor modification might make it even better: turn 96 can be the same as 97-100 rather than another D, since mutual D in 95 (after 94 mutual C) should be enough to establish clonehood. That way against clones it gets 99 mutual C and just 1 mutual D.

Also, if 1-94 aren't all cooperation, it might be better to have it just defect the last 2 or 3 turns rather than the last 4.

Comment author: 06 September 2011 08:59:24PM 0 points [-]

I said 95,96 are D,C. I should have said "respectively" to make it clearer.

Comment author: 07 September 2011 09:24:19PM *  1 point [-]

Actually, in the initial environment, using 96,97 for the unprovoked D,C (instead of 95, 96) will be more successful. Like O, but unlike the first afterparty, this "late afterparty" (LA) will beat I in an environment of mostly I (score against I for last 6 turns: I gets 18, afterparty gets 16, O gets 21, and LA gets 20). It will also beat O in an environment of mostly it and O (against LA, LA scores 21 and O scores 14; against O, LA scores 14 and O scores 15. So if the environment has more than 1/8 LA, then LA is winning. From the above, if O and LA start out small and equal in an I-dominated environment, as I disappears LA will be around 40%, easily enough to come back and pass O). LA will lose to O in a strongly O-dominated environment , while the original version will win (Afterparty scores 17 against O, above O's mirror-score of 15); but since that situation should never arise, I think that LA has the best chance of coming to dominate the environment. In fact, in an environment which started out I-favorable and with both O and both forms of afterparty, original afterparty's only effect would be to hasten LA's win by parasitizing O.

(Edited; I'd originally used the 3,5,0,1 payoff matrix instead of 4,7,0,1)

Comment author: 07 September 2011 03:02:24AM 0 points [-]

Hmmm, for some reason I'd read that as D,D.

I still think it makes more sense to group 96 in with 97-100, though, as a turn where it only plays C if all matches so far have agreed. If 1-94 are all mutual C, then on 95 it plays D and opponent plays C, I'd think it should play D the rest of the way instead of going with C on 96 and then D on 97-100.

Comment author: 07 September 2011 03:17:19AM 2 points [-]
Comment author: 07 September 2011 01:56:10PM 1 point [-]

A C on 96 can benefit if the opponent is still TFT on 97.

Basically, going to defect too early can drive you to extinction when the rest of the population is still late-defectors.

Comment author: 08 September 2011 06:06:09PM *  1 point [-]

I ran some numbers and it looks like your way is better against late-defectors.

Here are 3 ways you could design Afterparty:

v1. D-C on 95-96, then D 97-100 against non-clones (your original proposal)
v2. first D on 95, then D 96-100 against non-clones (my suggestion)
v3. first D on 96, then D 97-100 against non-clones (a variation on my suggestion)

My intuition was that v3 would be better than v1, even if v2 was not, but it turns out that's not the case. Version 1 scores the most against TFT (or variations of TFT that defect on up to the last 3 rounds). Compared to the others v1 has an extra CD and an extra DC (worth 7 points in total), v2 has two extra DD's (worth 2 points in total), and v3 has an extra CC and an extra DD (worth 5 points in total).

Things only change if the opponent is designed to play D for the last several rounds. If the opponent is TFT with D on the last 4 rounds then v3 does best and v1 does worst, and if the opponent plays D on the last 5 rounds then v2 does best and v3 does worst.

Comment author: 09 September 2011 12:19:11AM 1 point [-]

I already ran the numbers, and I think that Late Afterparty, version 4 with D-C on 96, 97, would have been most likely to have won this tournament.