For years, I've toyed with a similar idea to this. I feel like a lot of work on this has relied on some pretty crude heuristics like "rolls" and "alternation". I think it's dangerous to overgeneralize about these, and I think it's also short-sighted because there are probably more specific patterns that affect typing speed.

So my idea was to collect data on all keystroke transitions. There are 1,089 of these on the part of the keyboard that you would optimize for. Typing 5 minutes a day at a relaxed pace would be enough time to cover each of those 8 times in a month, so collecting enough data to work with would be feasible pretty quickly. I was then going to take a big textual corpus of English, and the total of all transition times would serve as the fitness function.

I had the data collection program done, and was starting to collect data, when I decided I should redo it with triples (of which there are 35,937). Then the length of the data collection process I was looking at became really daunting, and I lost interest in the project.

What I'd quite like to do, though, would be to crowdsource the data collection using a web app. Then I could get good three-key data and not have it biased to my own personal quirks.