New abstract, introduction, various minor revisions throughout, and s…

[marcharper]

…lightly modified title

I have some additional suggestions for the plots and tables that might help the main insights pop more directly.

Figure 2

• Plots need to be larger / split more. There's too much whitespace and the plots are too small.
• Consider dropping or moving some or all of the cooperation plots: SSE and \Chi are the main story.

Figure 3:

• Remove the best line IMO
• Maybe a heatmap is a better representation of the data

Meta comment (relevant to the caption in Figure 3 and throughout) -- it's not possible to guarantee an extortionate payout so there's a big difference between "attempting to extort one's opponnet" (e.g. acting extorationately) and actually extorting the opponent. For example, versus Defector, one cannot gain any advantage whatsoever. So when we claim e.g. that low SSE doesn't necessarily indicate extortionate behavior because \Chi is small, this doesn't mean that the strategy isn't trying to be extortionate. It just means that it's failing to be extortionate.

It's probably better to use e.g. Random(0.5) as a universal test strategy than the full basket of strategies in Axelrod when attempting to quantify if a given strategy attempts to be extortionate or not (rather than if it is able to achieve extortion).

Table 2 and Figure 4.

• Too busy IMO
• Takeaway message seems to be that mean score and variance of SSE are related. Would it make more sense to plot SSE as a function of tournament rank instead? One would expect to see an uptick on the high end where the evolved strategies increase their score a bit by exploiting the weaker opponents. Technically this information is in the plots but it's hard to see if you don't know to look for it.
• Similarly for SSE versus rank by wins (maybe could be on the same plot as SSE by score)
• Similarly for \Chi (mean plotted as a function of rank)?
• In general I prefer a targetted image with the main point rather than an "all the data" approach (in the main text).

Figure 5: Would it be better to simply plot the final population probabilities as a function of score rank?

Figure 6:

• I like this figure a lot
• The takeway I get is that while the mean doesn't seem to correlate with the long run proportion, the variance does; nevertheless there is a clear cluster in both plots.
• I would remove the lines (on the left plot, at least) and all references to R^2 since the data isn't linear: there's a line of strategies at zero and a cluster of other strategies above it, which seems to be the main insight, which is pretty clear.
• I removed the text about the predictive power being low. If it were linear that is true, but given the cluster and the line, one can imagine other models (like a SVM or just a linear separation) that would classify strategies.

Table 3: The model seems to say that high mean and low median give a lower s_i but the caption says the opposite. Am I interpreting the coefficients incorrectly?

Figure 7 / Table 4:

• There seems to be a similar clustering effect as in Figure 6 so I'm not sure a linear regression makes sense in the bottom left plot. You can imagine the bottom right plot converging to the right hand plot of Figure 6 (which it would for N \to \infty).
• The takeaway in terms of the mean and median are still the same. For comparison maybe it makes sense to leave in the variance in the model so that it has the same features at the model in Table 3.

[drvinceknight]

This all looks awesome @marcharper, I'm going to merge and open an issue to jot down some of my thoughts on your points :+1:, I think a chat will be very helpful though to go over some of them together. I'll also aim to start work on making the plots "nicer", it feels like we're getting to the point of being ready to do that. :+1:

[drvinceknight]

115