experiment: first impressions matter

[maxkw]

show two reciprocal agents playing each other. show that an 'error' earlier in an interaction is more costly i.e., will lead to more mutual defection for a time than an 'error' later in a sequence.

[polyguo]

So one should be mean a few times in a row at first? In an environment with no tremble this is just screwing with the observer, as they cannot conceive of someone doing something mean unless they're a bad guy. It will of course believe them to be bad and it will be costly. I've been thinking a lot about these "unnatural" occurrences.

What about having a decaying vs increasing tremble, I suspect that having an increasing tremble will still be better because during the early days reputations will be established and later defections can be explained away by agents. In the opposite case, the high-tremble early days are very uninformative, and people only come to know each other later in the game.

[maxkw]

Errors can happen for all sorts of reasons it could be tremble but it could also be noise in the softmax. Let's assume 0 tremble for now. I didn't mean a sequence of selfish behavior but rather just a single selfish event early on vs. later on. If its early on, it will really affect the belief leading to a higher likelihood of a tat-response. If it happens later on it will not affect the belief that much making cooperation more stable. Basically the goal is input a constructed sequence:

CCCCCCCD... CD.... and compare what happens afterwards.

Now we can see how adding tremble affects and/or modulates this effect.