Weirdly, it seems[0][1][2], that the layout doesn't do particularly well by Carpalx's keyboard effort estimator[3], as it gets a total effort score of 2.260, compared to 3, 2.098 and 1.842 for Qwerty, Dvorak and Colemak, respectively[4] (lower is better). In contrast, the halmak repo claims[5] an efficiency gain of +134% over Qwerty for halmak, compared to +77% for Dvorak and +84% for Colemak. Since (one would assume) efficiency and effort should be approximately inversely correlated, these results don't square.
I have no idea whether Carpalx's or Halmak's keyboard layout (efficiency|effort) evaluator is more accurate (indicative of real (efficiency|effort)), but the huge disparity between the two is troubling.
(Obviously, Carpalx and Halmak use different optimisation methods — Monte Carlo simulated annealing versus an "evolution algorithm based AI" — and these could well give different results, even with the same scoring method, due to getting stuck in different local maxima etc., but the issue is that the scoring methods themselves give different rankings of fixed keyboard layouts.)
Edit: quickly reading through all of the halmak author's blog posts, it seems that they're aware of the Carpalx project. They describe their (current?) effort model here[6]. It seems heavily data-based, but it's still difficult to determine whether it's "better" than Carpalx's.
I have no idea whether Carpalx's or Halmak's keyboard layout (efficiency|effort) evaluator is more accurate (indicative of real (efficiency|effort)), but the huge disparity between the two is troubling.
(Obviously, Carpalx and Halmak use different optimisation methods — Monte Carlo simulated annealing versus an "evolution algorithm based AI" — and these could well give different results, even with the same scoring method, due to getting stuck in different local maxima etc., but the issue is that the scoring methods themselves give different rankings of fixed keyboard layouts.)
Edit: quickly reading through all of the halmak author's blog posts, it seems that they're aware of the Carpalx project. They describe their (current?) effort model here[6]. It seems heavily data-based, but it's still difficult to determine whether it's "better" than Carpalx's.
[0] https://gist.github.com/tdegrunt/80e63f464c9a1c336e0f1d4e6aa...
[1] https://github.com/MadRabbit/halmak/issues/4#issuecomment-44...
[2] I haven't done the analysis myself and can't be sure that it was done correctly
[3] http://mkweb.bcgsc.ca/carpalx/?interpreting_optimization
[4] http://mkweb.bcgsc.ca/carpalx/?keyboard_layouts
[5] https://github.com/MadRabbit/halmak#comparisons
[6] http://nikolay.rocks/2016-10-22-keyboard-analytics