How does the AREDN firmware select RF link rates?

Here is the best consolidated references of documentation of the link rate selection I've found:
http://blog.cerowrt.org/post/minstrel/

The original paper was never published.  

Relevant forum post: http://www.arednmesh.org/comment/5321#comment-5321

Ironically, there was a post today in the wireless ath9k groups (the driver we use in AREDN firmware) about exploring better techniques to improve the tradeoff between packet error rates and latency (impacted by retries).  Traditional thinking is that the RF modulation rates are better selected with 0% error rates.  But there are those that suspect some error rates may be acceptable to achieve better overall performance.  One guy is investigating rate selection "based on Bayesian statistical inference, possibly by modeling the rate control problem as a "multi-armed bandit" problem and/or using Thompson sampling".

This stuff gets into the depths of statistical modeling and heavy math to engineer the best fit to optimize the problem.  Pull out your calculus and statistical math books to proceed :) .

Joe AE6XE