GuySrinivasan comments on Fixed-Length Selective Iterative Prisoner's Dilemma Mechanics - Less Wrong

25 Post author: Andreas_Giger 13 September 2011 03:24AM

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

Comments (14)

You are viewing a single comment's thread.

Comment author: GuySrinivasan 13 September 2011 08:35:09PM 3 points [-]

I sketched an entry for the original contest while on vacation, but by the time I went to write it up (before Sept. 1!) submissions were closed. My sketch looks very similar to 18-87/TFT-2D, and (as I recall) the relevant parts of the reasoning that led me to it were:

  • a TFT variant will win, so I'll write a TFT variant
  • clearly the winner needs to beat TFT so I'll defect on the last turn
  • most people will defect on the last turn so maybe I'll defect on the last two turns
  • (I don't care about the elimination thing, only the evolutionary tournament)
  • given that I'm around in the last few generations, I will compete against me, so I'd like to not defect against myself in the last two turns
  • I'll pick some arbitrary turn to defect on like 56, and if I see that we both defected on 56 then I'll cooperate for the last two turns
  • ...but 1/1,4/4 is fewer points than 0/7,7/0 so it'd be better if I could defect on different turns against myself
  • fine, randomly between 50-80, and on turns 99-100 if my opponent acted like TFT except for a single defection, then I'll cooperate. Or okay, why not 1-95. No, people will probably do silly thing in the first few turns. And last few. 18-87 or something.

I'd like to see competitors' reasoning for their submissions.

Comment author: Andreas_Giger 14 September 2011 05:53:42AM 1 point [-]

Looks reasonable. However, point gain from increasing the range of i is logarithmic while the danger of opponents randomly (from your perspective) defecting within that range increases approximately linearly, so large ranges aren't optimal. I only used 1-∞/TFT-nD as an example so I wouldn't have to deal with decimal numbers.

Comment author: FAWS 13 September 2011 10:57:12PM *  0 points [-]

My reasoning was pretty much exactly the same (only diverging slightly on the last bullet point), but I didn't do well due to the presence of vengence bots.