I’d be very interested in seeing the same sort of thing where the number of rounds is randomly chosen and they don’t know how many rounds they are going to go up against.
Note that the CliqueBots failure is not a good reason to think that Cliquebotish like strategies will fail in evolutionary contexts in real life. In particular, life generally has a lot of conditions to tell how closely related something is to them (such as skin coloration). In this sort of context cliquebots are losing in part because they need at one point to defect against themselves to do an identity test. I suspect that if strategies were told in advance whether they are going up against themselves then Cliquebots would do a lot better.
P seems to be a much better designed CliqueBot than Q especially for large sets. Five defections is a lot to burn, and in fact it does go away well before Q leaves.
One other thing that might be interesting to see what sort of environments which strategies will do well in is to never reduce the number of copies of any strategy to 0, and have 1 be the minimum.
One other thing that might be interesting to see what sort of environments which strategies will do well in is to never reduce the number of copies of any strategy to 0, and have 1 be the minimum.
I think CliqueBots are still going to do badly if there’s only one member of the clique :)
I suspect that if strategies were told in advance whether they are going up against themselves then Cliquebots would do a lot better.
It might also be interesting to see what would happen if the bots were allowed to either remember their last N scores and/or win/loss record for the last N games, or know how many other bots of their type exist in the population. I suspect that “if I’m a member of a small minority do X, else if I’m a member of a large minority or small majority do Y, else do Z” would be an interesting class of strategies to look at.
P is better than Q not because Q’s initial five defections, but because it plays more than half of the match like TfT. So it’s basically only a 43% CliqueBot. The problem with CliqueBots (under rules of this game) aren’t resources spent testing the opponent’s identity, but that they are cooperating neither with the nice strategies nor with CliqueBots of other types.
I think communication cost isn’t main reason for P’s failure. O, for example, defects on 3 last turns even when playing against itself (rule 1 is of highest priority). Reasons are too strong punishment of other strategies (and consequently itself) and too strict check for self identity.
Strategy I described here should perform much better, when n is in range 80-95.
I’d be very interested in seeing the same sort of thing where the number of rounds is randomly chosen and they don’t know how many rounds they are going to go up against.
Note that the CliqueBots failure is not a good reason to think that Cliquebotish like strategies will fail in evolutionary contexts in real life. In particular, life generally has a lot of conditions to tell how closely related something is to them (such as skin coloration). In this sort of context cliquebots are losing in part because they need at one point to defect against themselves to do an identity test. I suspect that if strategies were told in advance whether they are going up against themselves then Cliquebots would do a lot better.
P seems to be a much better designed CliqueBot than Q especially for large sets. Five defections is a lot to burn, and in fact it does go away well before Q leaves.
One other thing that might be interesting to see what sort of environments which strategies will do well in is to never reduce the number of copies of any strategy to 0, and have 1 be the minimum.
I think CliqueBots are still going to do badly if there’s only one member of the clique :)
It might also be interesting to see what would happen if the bots were allowed to either remember their last N scores and/or win/loss record for the last N games, or know how many other bots of their type exist in the population. I suspect that “if I’m a member of a small minority do X, else if I’m a member of a large minority or small majority do Y, else do Z” would be an interesting class of strategies to look at.
P is better than Q not because Q’s initial five defections, but because it plays more than half of the match like TfT. So it’s basically only a 43% CliqueBot. The problem with CliqueBots (under rules of this game) aren’t resources spent testing the opponent’s identity, but that they are cooperating neither with the nice strategies nor with CliqueBots of other types.
I think communication cost isn’t main reason for P’s failure. O, for example, defects on 3 last turns even when playing against itself (rule 1 is of highest priority). Reasons are too strong punishment of other strategies (and consequently itself) and too strict check for self identity.
Strategy I described here should perform much better, when n is in range 80-95.