• You are not logged in.

Best Layouts for Hand Alternation?

  • Started by lalop
  • 41 Replies:
  • Reputation: 114
  • From: Oslo, Norway
  • Registered: 13-Dec-2006
  • Posts: 4,742

Imagine trying to learn a new layout with, say, 105 haphazard key placements. Everything jumbled. Just as a thought experiment. I think it'd matter.

The general idea in the debate that led to Tarmak being made was that any change will disrupt the flow and "gestalt feeling" of your typing, but the complexity of the changes will influence how fast you get back into the flow. Many Tarmak users have attested to this. Maybe you're different though, who knows. But intuitively it should matter how complex a change is!

Last edited by DreymaR (27-Jul-2014 10:47:24)

*** Learn Colemak in 2–5 steps with Tarmak! ***
*** Check out my Big Bag of Keyboard Tricks for Win/Linux/TMK... ***

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136
DreymaR said:

Maybe you're different though, who knows. But intuitively it should matter how complex a change is!

I am learning the hard way that complex changes ARE difficult. Why am I trying it anyway ?
1. I am a stubborn extremist. If you optimize, why not go all the way?
2. I never learned to touchtype properly, I hunt&pecked. So, I am new to touchtyping. A clean slate..... so why not learn the complex layout?

This being said - man, it is hard. I am learning 2 things at the same time: touchtype and a radically different layout.....So if you are wiser and less stubborn than me, you better learn Colemak !

OTOH, people are able to learn the Dvorak layout, my layout is not harder....

Offline
  • 0
  • Reputation: 0
  • Registered: 04-Apr-2013
  • Posts: 538

Continued from https://forum.colemak.com/viewtopic.php … 662#p15662

lalop said:
pieter said:

Dvorak favors left-right, left-right (LRLR etc), my layout favors LLRRLLRR etc.

I'm wondering how you determined this (I don't think my own analyses ever were able to detect such things).

pieter said:

@lalop: for my own Juli16 layout:
a) based on feel. Admitted, this is subjective... 
b) based on the objectives of the mtgap model, as michael dickens writes (again, I admit this is a weak argument)
c) based on hand alternation stats. Dvorak has a same hand stat of 31% (see above).  I interpret this (correct me if I'm wrong) that for any given key stroke, the chance that the next stroke is on the same hand is 31%. LLRRLLRR would indicate that on average, 50% of all keystrokes are same-hand.
LRLRLRLR would have 0% samehand. LRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLRLLRRLRLRLRLRL are 70 key changes. 68 of those switch to the other hand, 2 stay on the same hand. This would imply a same hand% of 3%  (right? )

I now see that I am wrong in two ways on this stat. In the first place because LRLRL would mean 0% same hand. My layout scores 36% same hand. which means roughly that 1 in three strokes is on the same hand. LLRRLLRR would be 50%. My second mistake is that this is percentage over all key strokes in the corpus. It's an average, and I am also interested in the spread (variance etc.) 

For Dvorak:
a) based on Dvoraks ideal to have high alternation, by putting all vowels on one hand. A typical word in languages like English, Dutch, French and German consist of a consonant, a vowel, a consonant (CVC). Look at the words typical or German. Of course there are words like area (VCVV). Spanish has more words of a VCV structure (ave, olá, odio etc.), German has more consonants (schmerz, mehrfach), slavonic languages have even more (like the chech "Krkonošský národní park"), more like CCCVCCVCC :-)
b) based on subjective feel. 

But, all in all I admit to jump to conclusions here. This screams for a Pyhton script to analyse a corpus. The script would need three inputs: lefthandlettersoflayoutX.txt and righthandlettersoflayoutX.txt plus corpus.txt. It would give two outputs: one LRLLRRLL text file, and one file with stats..... Let's see if I can find some time....

lalop said:

LLRLLRLLRLLRLLRLLR is an example of 33% same-hand, and yeah, LLRRLLRR would be 50%.  (On the other hand, so are LLLLLLLRLRLRLRLRL and LLLLLLLLLLRLRLRLRL respectively.  This is another reason same-hand percentage is not a very insightful measure.)

A good way to make a general analyzer might be to take patterns (e.g. LLRR, RRLL vs LRLR, RLRL) as input and just count the number of occurrences.  I'm not 100% sure we'd be able to interpret this data as "one tends toward LLRRLLRR and the other toward LRLRLRLR", but it'd be a start.

I haven't had time to update any of my analyzers, unfortunately.  Feel free to try your own hand at one.

pieter said:

Regarding the analysis, indeed: counting LRLR etc patterns, and calculating percentage perhaps? I'll chew on this!
I was thinking of a different solution: giving letters a value. The value of the bold letter is calculated as a function of the letters that surround it (I ignore the first and last letter of the text for now. Also: spaces don't count) 

LRL -> 2  (R is surrounded by 2 L letters)
LRR -> 1 (R is surrounded by 1 L letter)
RRL -> 1 (R is surrounded by 1 L letter)
RRR -> 0 (R is surrounded by 0 L letters)

same thing for the Ls. 

Next, sum the numbers and divide by the total number of letters in the text. In a 1x1 text, every letter will have value 2. So, the mean value (total/#letters) is 2.  So, 2 is maximum alternation.

In a 1x0 text (all Ls or all Rs), all letters will have the value 0. Mean is 0 as well. So 0 = no alternation at all.

In a 2x2 text, every letter will have an L and an R as neigbour, so every letter has value 1. Mean is 1 = 2x2.

Let's take a 3x3 text.  LLLRRRLLLRRRLLLRRRLLLRRR. Ignoring the starting and finishing letter, the values are 011011011011011 etc. Mean value is (0+1+1)/3 = 0,67.

Finally, 3x1x1x2. This one is irregular (no fixed rithm) and imbalanced: the left hand has 3 + 1 = 4 strokes, the right hand has 1 +2 = 3 strokes.  Again, ignoring first and last one.  LLLRLRRLLLRLRRLLLRLRRLLLRLRR  This gives 012211121221 1012210121221 = 32 / 28 = 1,14

I'm not happy with this yet, because irregularity does not show in the metric..... I'll think some more.

An alternative metric might be to count how many letters one has to travel to the left/right before reaching an alternation.  In this case,

RLRLRLRL
00000000 -> 0/8 = 0

RRLLRRLLRRLL
1111111111111 -> 13/13 =1

LLLRLRRLLLRLRRLLLRLRRLLLRLRR
2220011222001122200112220011 -> 32/28

Implementation details: This "starts counting from 0" (so counting rightwards from the first R in RL would get 0 rather than 1, and from RLLR would get 2 rather than 3) and assumes the endpoints are alternations.

One could also "weight" various distances. Since we care about larger distances more than smaller ones, perhaps try a weight function like w(dist) = dist^2.  Then the first two stay the same and the third

LLLRLRRLLLRLRRLLLRLRRLLLRLRR
4240011424001142400114240011 -> 48/28

Different people might use their own weight functions, based on what distances they care about.


With the above examples, I count bidirectionally.  That's probably unnecessary/a bad idea, though.  Counting in only the rightward direction and multiplying the end-score by 2 should give the same result, aside from the left endpoint (if one starts adding from an "invisible left endpoint" of the opposite hand, it should give exactly the same result).  Of course, this also begs the question of why multiply by 2 at all (aside from it giving cleaner numbers for the RLRLRLRL and RRLLRRLL cases).

Last edited by lalop (13-Sep-2014 04:24:59)
Offline
  • 0
  • Reputation: 114
  • From: Oslo, Norway
  • Registered: 13-Dec-2006
  • Posts: 4,742

You should never count bidirectionally. Backwards typing is satanic. ;)

*** Learn Colemak in 2–5 steps with Tarmak! ***
*** Check out my Big Bag of Keyboard Tricks for Win/Linux/TMK... ***

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

@Dreymar: LOL
@lalop - good remark about making some sort of statistic. I'll dive into statistics, I think what we have here is a special case of comparison of distributions. Time to look for an appropriate test. Or, the other way around, to recalculate the problem so that we can use e.g. a Chi-square test..... The test would then be: how closely does this alternation fit a 1x1 (or 2x2 or 3x3) alternation....

Edit: this might be useful: Damerau-Levenshtein metric for string distance

Last edited by pieter (13-Sep-2014 21:33:25)
Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

@ lalop, I understand your distance metric, but maybe an easy solution is good enough? (yes, I dropped even the Damerau-Levenshtein metric :-) Say, a text has 10,000 letters and you want a perfect 1x1 alternation  (LRLRLRLRLR). You could fit the digram LR maximum 5,000 times in that text (actually, 4,999 times. Let's just call that 5,000  OK?) And by the way, i only count from left to right :-)   

So, simply count the number of LR digrams in the text. Let's say, there are only 3,000 instances of LR in the text. In that case, the test for "is this a 1x1 layout?" is: 
- calculate predicted # of LR digrams: 5,000
- count actual # LR digrams: 3,000
- divide actual/predicted =  3,000/5,000 = 0.6 = 60%.  In other words: the layout has a 60% fit. That is, not a very good fit. In other words: it is not a good alternating 1x1 layout. 
(note: in reality, 100% scores are unlikely. Actually, I do not know what are acceptable percentages. in other words, I don't know what would be a good cutoff point)

Now for LLRR (2x2) layouts. The same thing. In a 10,000 letter text, you can expect maximum 2,500 instances of this 4-gram. The test is:
- calculate predicted # of LR digrams: 2,500
- count actual # LR digrams, let's say we find 1,000
- divide actual/predicted =  1,000/2,500 = 0.4 = 40%.  In other words: this layout has a 40% fit. Not a good alternating 2x2 layout. 

Something along this line, lalop? Is this going anywhere?

Last edited by pieter (14-Sep-2014 19:23:46)
Offline
  • 0
  • Reputation: 0
  • Registered: 04-Apr-2013
  • Posts: 538

Hard to say; why not both?

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

Update on alternations.  i had an interesting e-mail conversation with martin Krzywinski from Carpalx, who wrote this specific Dutch language layout for online magazine De Correspondent:

QFKLMWVUJY[]\
SDTNRIAEOG'
;ZBHXPC,./

I constrasted it with my own Juli16 layout:
. u o p y x c l b v
a i e n h m d r t s
: , ? k q f g w j z

Martin showed me good statistics that indicated that mine is indeed rather left hand sided, plus can have long slurs of letters, specifically on the left hand. The main reason is that ENA are the most used letters in Dutch, and EN is by far the most used digraph. In is also on that hand, IE is a much used digraph and AN as well. So, there are more strings of 3, or 4 letters on the left hand; than there are on the right hand. This leads to an unbalanced feel  (very good nevertheless in terms of distance)   

This got me thinking again, and I played with MTGAP's settings and got this:

. u o f y q p l m w
a i e s h c d n t r
: , ? k x b g v z j

This one is much better balanced (52/48, versus 57/43) and has shorter strings. Better alternation. But, at the cost of some 'breakage of nice rolls'  such as EN, DR and ST, which have disappeared here. According to Patorjk's stat "consequetive hand use" as good as Dvorak. I will try this one out and also that Carpalx one, as an aside, I see some less nice features in that one - carpalx assumes that all positions on a row are equally good. Which they are not. So some letters are in too good positions and some in too bad spot, like the W that he puts in the dreaded qwerty Y spot. I also don't like that R and I are not under the fingers on the home row - they deserve that!  

TL;DR

1. carpalx has nice, very nice, detailed stats on alternation. Look at his program
2. patorjk also has some stats (i dont know how he calucates it though)
3. with mtgap you can play with the settings and find very alternating layouts
4. alternation goes at at cost of other things.  aiesh has more alternation and better balance than aienh (a.k.a. Juli16), but at the cost of rolls.
5. you (lalop) were looking for corpus text,right? krzywinski has published freqs, digraphs and trigraphs on the Dutch language, based on a very large corpus, much larger than mine. see the carpalx site, under ' correspondent'

Last edited by pieter (29-Sep-2014 11:25:48)
Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136
lalop said:
pieter said:

Dvorak favors left-right, left-right (LRLR etc), my layout favors LLRRLLRR etc.

I'm wondering how you determined this (I don't think my own analyses ever were able to detect such things).

By hindsight, you (Lalop) were right in your criticism, I had no real metric to determine. Now that I have seen how carpalx calculates them, I know I was wrong. My AIENH layout is 57% lefthand. Worse is that there are more longer strings on the left hand than on the right hand. I have also found out (for some this may be trivial) that it is hard to have both rolls and good alternation. Look at the desirable characteristics of layouts, in no particular order:

LOAD
a- low 'same finger'
b- low row switching
c- low home jump
d- some pre-set balance of fingers load (e.g. 9/9/17/15)
e- balanced hand load (50/50)

DISTANCE
a- low total distance

RHYTHM
a- good alternation Left - Right 
b- rolls (inward rolls, outward rolls, etc)


It's very hard to find a layout that does well in all. My guess is that the rhythm metrics are the hardest. Specifically, the two rhythm variables seem to be mutually exclusive.

Last edited by pieter (29-Sep-2014 19:27:04)
Offline
  • 0
  • Reputation: 0
  • Registered: 27-Jul-2014
  • Posts: 35

I actually find the ring to pinky roll easier to do than the pinky to ring roll on the right hand, with adapting to the ion trigram. I don't really like it when it roles the other way, like in going. I tend to pause a little after the I key trying not to accidentally hit the E key as well, But I don't really have that pause problem with other words that have the "ing" trigram in them without the letter before that trigram being a "O".

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

What about this theory: hand alternation is better for languages like Dutch, German or Finnish;  rolls for Spanish.  English is in between.

Explanation: Dutch, German and Finnish have many words with repeated letters. Like brooddeeg, schwimmmeister, loop, kook, meer, rennen and so forth. This is because of 2 things:
- a letter can be duplicated to change the sound (man  versus maan)
- words can be glued together. English would say "bomb threat " but Dutch would say "bommelding" (bom meaning bomb; melding meaning threat) Glueing words toegether means more repeat letters, in this case the m

When you have many repeated letters, you cannot avoid using the same finger (and thus the same hand). Adding rolls on top of that would overload the hands. A highly alternating layout may be better.

In Spanish, by contrast,  letters are rarely repeated. Only the rr (like in autorretrato, correa, horroroso and so on), the ll (like in llama, paella) and the cc (dirección, acción and so on). And the very rare double nn like in innato (in born). But mostly, it is a clean 1 letter system: "el hombre y la mujer hablan de sistemas y teclados"  No repeat keys. So there is less same finger load, which means that an alternating scheme may become TOO alternating and that rolls may be better.

Thoughts?

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

@davkol -  great idea. Some people on the Google ADNW-group mentioned though that a repeat key might be complicated to learn. Should try it out, actually. I had also been thinking about mapping the most used digrams to keys. That would solve a part of the problem (of too high same hand strings) as well.

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

Yes. Part of it comes from glueing words together, but the rest comes from letter repeating to make the sound longer, or a bit different. Indeed, like diacritics. I never thought of it that way but you are right! The question is, could you get used to spelling "letter" L E T repeat E R  or (like you suggest) L E deadkey T E R.

It may be too much "cognitive overhead". I don't know. If so, maybe an other idea is easier to get used to, namely: having keys for much used digrams or even trigrams. Like (in English) TH IN RE. However, this gets into steno or Auto(hot)key territory.

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

Comning back to the original poster's question: what are the best layouts for hand alternation? Recently I have worked with ADNW and the ADNW optimising software, which is great. The support is in the ADNW Google Groups. One possible disadvantage for some: it is mostly in German, so you have to make translations if you cannot read that. 

The algorithm can be tweaked in many ways, both by modifying the source code (it is in C++) and at runtime, by giving flags or modifying settings (which are in .txt files). ADNW is aimed at maximizing alternation and minimizing rolls of neigbouring fingers. The highest "penalties" are for outward rolls and for ring/pinky. So (qwerty) sa l; are worst. (qwerty) sf lj are the best rolls. The program can analyse digrams and trigrams, in trigram setting "too much" alternation is penalised as well, so (qwerty) SLAIDLE would be considered as worse then (qwerty) SDKJDF.  (disclaimer: as far as I understand it !!)

Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

I calculated lots of keyboards with the adnw algorithm. Compared to the standard settings, I used slightly higher penalties for the central keys (qwerty ghy) and higher penalties for outward rolls, because adnw in stock settings produces to outward rolling layouts in Dutch (meaning it optimized too much on other factors).

I translated the words in the outputfile from German to English, for your comfort :-)

Totalscore: the total score of the layout (lower is better). This is the base score, plus penalties for same finger, etc.
Basescore: the base score of the layout, I think this is comparable to Carpalx base path and Patorjk's distance. Lower is better.
left right: percentual division of hands. Normaly, 50% / 50% is ideal. But of course you might go for a different distribution
Same Finger: the percentage of bigrams typed with the same finger. Lower is better. There are is an extra penalties for row jumps
Shift-Same Finger: for pinkies: letter followed by shift (I think!) , or vice versa. Lower is better.
Indirect same finger: I don't know! I think it is like qwerty FOR, in which the F and the R are type with the same finger. Lower is better.
up, mi, bo: percentage of keys hit on the upper row, mid row and bottom row
Alternation: percentage of bigrams that are typed with two hands. More is better. However, in trigram optimization, there is a penalty for too much alternation. A trigram like (qwerty) DOS is left- right-left, which is too much of a good thing. 
Shift-Alternation: I don't know!
In/outwards: the relation inward moving bigrams versus outward moving bigrams. More is better.
Neighbouring: the percentage of bigrams that are typed with two neighbouring fingers. Especially pinky-ring is penalised in ADNW. The assumption is that it's easier to type (qwerty) SF than SD.       Shift-Neighbouring: ?
Below is distribution of keystrokes over the fingers, from left pinky to right pinky. 
(BTW, I left the äßöü in the keyboard, those are not used in Dutch or English, I'll put stuff like ? / " ! on it and I like to optimize those by hand.)


You see that my layout (I nicknamed it Nordsee, German for Noth Sea) improves on Dvorak on same finger, Inwards motion, neigbouring fingers . Alternation is about the same. The total score of North Sea is much better.

ADNW is calculated on 50% German/ 50% English. I calculated Nordsee on a 75% Dutch, 25% English corpus.



DUTCH TEXT 

Nordsee            274.376 Totalscore  162.896 Basescore                left right
                   1.288 Same Finger        4.009 Shift-Same Finger    up 9.4  7.8
  p.ouä xclhzq    68.057 Alternation       18.652 Shift-Alternion     mi 39.5 30.5
  tiaej gdrnsk     1.413 In/outwards        3.792 IndirSameF           bo 2.8  9.8
  ö,üyß vwmfb      6.998 Neighbouring      19.091 Shift-Neighbouring sum 51.9 48.1
                  9.9  8.9 12.1 21.0 --.- --.- 14.1 11.8 12.3  9.8    Sh  2.0  1.0

For comparison:
ADNWstock           326.733 Totalscore  178.160 Basescore          left right
                   1.839 Same Finger      2.894 Shift-Same Finger  up  4.9 12.8
  kuü.ä vgcljf    70.890 Alternation     29.801 Shift-Alternation  mi 39.0 31.3
  hieao dtrnsß     1.988 In/outwards      7.172 IndirSameF         bo  2.8  9.3
  xyö,q bpwmz      9.571 Neighbouring    19.504 Shift-Neighbouringsum 46.6 53.4
                   6.0  8.7 18.0 13.9 --.- --.- 20.4  9.3 15.6  8.0 Sh  1.9  1.1

Dvorak            388.597 Totalscore  201.815 Basescore        left right
                   2.962 Same Finger     11.975 Shift-Same Finger     up  3.4 15.1
  ä,.py fgcrlq    69.288 Alternation     25.341 Shift-Alternation     mi 38.7 27.0
  aoeui dhtnsß     1.256 In/outwards     8.885 IndirSameF             bo  5.7 10.1
  öüjkx bmwvz     12.039 Neighbouring   19.404 Shift-Neighbouring    sum 47.8 52.2
                  8.6  6.1 20.5 12.6 --.- --.- 14.7  9.6 18.5  9.4    Sh  2.0  1.1

Now let's see how English text does on this keyboard. I didn't translate the german words in the output, but you'll get it now.

ENGLISH TEXT 

      (Nordsee)  332.473 Gesamtaufwand  189.723 Lageaufwand        links rechts
                   1.492 Kollisionen      3.609 Shift-Kollisionen  ob 12.8 12.1
  p.ouä xclhzq    70.575 Handwechsel     24.160 Shift-Handwechsel  mi 35.3 25.6
  tiaej gdrnsk     1.157 Ein-/Auswärts    5.961 IndirKollision     un  4.4  9.9
  ö,üyß vwmfb     10.208 benachbart      25.369 Shift-benachbart  sum 52.4 47.6
                 12.1  9.4 14.9 16.0 --.- --.- 11.6 12.3 13.7 10.0 Sh  1.6  1.2

ADNW    316.834 Gesamtaufwand  182.526 Lageaufwand        links rechts
                   1.020 Kollisionen     17.224 Shift-Kollisionen  ob  4.5 12.0
  kuü.ä vgcljf    71.389 Handwechsel     21.757 Shift-Handwechsel  mi 38.3 31.6
  hieao dtrnsß     2.136 Ein-/Auswärts    7.795 IndirKollision     un  5.0  8.6
  xyö,q bpwmz      8.122 benachbart      20.009 Shift-benachbart  sum 47.8 52.2
                  7.6 11.2 11.7 17.4 --.- --.- 18.7 10.7 13.1  9.7 Sh  1.9  0.9

Dvorak Typ 1a    363.771 Gesamtaufwand  202.322 Lageaufwand        links rechts
                   2.668 Kollisionen     13.145 Shift-Kollisionen  ob  6.0 17.0
  ä,.py fgcrlq    70.679 Handwechsel     33.265 Shift-Handwechsel  mi 36.1 30.5
  aoeui dhtnsß     1.606 Ein-/Auswärts    6.623 IndirKollision     un  2.9  7.6
  öüjkx bmwvz     11.100 benachbart      20.835 Shift-benachbart  sum 45.0 55.0
                  9.7  8.2 13.0 14.1 --.- --.- 16.5 13.3 13.7 11.5 Sh  1.8  0.9

For English, ADNW has the best score - as expected. Look at those nice alternation figures (71%), very nice inward motion (inwards/outwards = 2.1) and low same finger rate (1%). It beats Dvorak in all metrics. If you only type English, a 100% English board will be stunning - that is, IF you like the Dvorak approach, which ADNW improved upon: high alternation, inward motion, low same finger rate, low neighbouring finger use. Mind that Colemak has different goals: it does not avoid by, instead, WANTS neighbouring fingers, it is not looking for alternation but for rolls. 

Remarkable is that Nordsee still beats Dvorak in most aspects, but of course it falls behind Adnw.

I will later calculate an alternating English only layout, for those interested.

Last edited by pieter (19-Nov-2014 15:51:31)
Offline
  • 0
  • Reputation: 2
  • Registered: 25-Oct-2013
  • Posts: 136

And here it is for English. Calculated only the  English corpus (so optimized for the English language). Since ADNW is sort of " Dvorak on steroids" I called this one DvorMax.

For comparison, I calculated how other keyboard layouts score with this same English corpus. I have grouped them into
-  a "Dvorak-like" category (DvorMax, Dvorak, MTGAP, Aus der Neo Welt (=ADNW, actually the same as DvorMax, but developed for 50% English and 50% German)  and Klausler),
- a "Colemak-like" category (Colemak, Capewell, Workman),
-  the "Mixed approach group"  (MTGAP).   
- and then there is Qwerty of course, for giggles. 

As always, the scores depend heavily on the algorithm used, in this case the ADNW layout.  Also, put you favorite characters (/;"?!and so on) on the spots äßöü.  The black box below contains many layouts, just scroll down in it! 

Scores of layouts. Lower is better.  ranked from best to worst 


CONTENDERS IN THE DVORAK-LIKE GROUP

DvorMax          291.491 Gesamtaufwand  179.388 Lageaufwand        links rechts
                   0.882 Kollisionen      7.423 Shift-Kollisionen  ob  6.1 13.1
  kyu.ö zlmdpv    65.718 Handwechsel     28.070 Shift-Handwechsel  mi 39.5 31.4
  rieao hnstcw     1.016 Ein-/Auswärts    6.177 IndirKollision     un  3.3  6.5
  xßü,ä jqfgb      8.665 benachbart       9.226 Shift-benachbart  sum 49.0 51.0
                  8.8  8.7 14.2 17.3 --.- --.- 15.8 11.0 14.3  9.9 Sh  2.0  0.8

Aus der Neo-Welt 337.581 Gesamtaufwand  189.071 Lageaufwand        links rechts
                   2.052 Kollisionen     17.995 Shift-Kollisionen  ob  4.5 12.0
  kuü.ä vgcljf    70.173 Handwechsel     25.534 Shift-Handwechsel  mi 38.3 31.8
  hieao dtrnsx     2.043 Ein-/Auswärts    7.458 IndirKollision     un  6.1  7.3
  yö,qb pwmzß      7.827 benachbart      16.064 Shift-benachbart  sum 48.9 51.1
                  8.9  9.6 12.9 17.6 --.- --.- 18.9 11.4 10.8 10.0 Sh  1.7  1.0

Dvorak           348.893 Gesamtaufwand  186.358 Lageaufwand        links rechts
                   2.662 Kollisionen     13.058 Shift-Kollisionen  ob  6.0 16.8
  ß,.py fgcrlö    70.517 Handwechsel     33.563 Shift-Handwechsel  mi 36.1 30.5
  aoeiu dhtnsü     1.617 Ein-/Auswärts    6.624 IndirKollision     un  3.0  7.6
  äqjkx bmwvz     11.139 benachbart      20.861 Shift-benachbart  sum 45.1 54.9
                  9.7  8.3 13.0 14.1 --.- --.- 16.5 13.3 13.7 11.4 Sh  1.8  0.9

Klausler         366.072 Gesamtaufwand  184.919 Lageaufwand        links rechts
                   2.506 Kollisionen     14.113 Shift-Kollisionen  ob  8.1 13.2
  k,uyp wlmfcß    68.668 Handwechsel     34.314 Shift-Handwechsel  mi 37.3 32.6
  oaeid rnthsü     0.678 Ein-/Auswärts    8.528 IndirKollision     un  3.2  5.6
  q.öäz xvgbj     14.512 benachbart      14.371 Shift-benachbart  sum 48.6 51.4
                  9.6 10.2 14.2 14.5 --.- --.- 19.5 13.0  8.3 10.6 Sh  1.8  1.0


CONTENDERS IN THE COLEMAK-LIKE GROUP

Colemak          355.884 Gesamtaufwand  185.391 Lageaufwand        links rechts
                   1.348 Kollisionen     14.374 Shift-Kollisionen  ob  7.8  8.2
  qwfpg jluyäö    58.143 Handwechsel     39.607 Shift-Handwechsel  mi 32.7 37.3
  arstd hneioü     1.041 Ein-/Auswärts    8.772 IndirKollision     un  6.8  7.3
  zxcvb km,.ß     17.492 benachbart       9.009 Shift-benachbart  sum 47.2 52.8
                  9.1  7.8 11.6 18.7 --.- --.- 18.8 15.4  9.8  8.7 Sh  1.1  1.7

Capewell         384.614 Gesamtaufwand  184.724 Lageaufwand        links rechts
                   1.779 Kollisionen     16.490 Shift-Kollisionen  ob 10.4  8.7
  .ywdf jpluqß    49.080 Handwechsel     51.941 Shift-Handwechsel  mi 33.8 30.9
  aersg btnioü     0.914 Ein-/Auswärts    6.664 IndirKollision     un  5.8 10.4
  xzcvä kmh,ö     21.958 benachbart      13.577 Shift-benachbart  sum 50.1 49.9
                 10.8 13.4 10.7 15.2 --.- --.- 15.3 15.5 10.8  8.4 Sh  1.6  1.2

Workman          392.782 Gesamtaufwand  191.440 Lageaufwand        links rechts
                   3.005 Kollisionen      8.626 Shift-Kollisionen  ob 13.0  6.8
  qdrwb jfupßö    55.794 Handwechsel     51.567 Shift-Handwechsel  mi 29.6 34.1
  ashtg yneoiä     1.088 Ein-/Auswärts    8.415 IndirKollision     un  7.6  8.8
  zxmcv kl,.ü     18.394 benachbart      12.531 Shift-benachbart  sum 50.3 49.7
                  9.0 10.5 13.2 17.6 --.- --.- 15.3 15.4 10.2  8.9 Sh  0.9  1.8


CONTENDER IN THE MIXED APPROACH GROUP

MTGAP            345.128 Gesamtaufwand  180.198 Lageaufwand        links rechts
                   1.270 Kollisionen      9.038 Shift-Kollisionen  ob 13.4 13.1
  ypouj kdlcwü    66.034 Handwechsel     40.419 Shift-Handwechsel  mi 34.5 28.4
  inea, mhtsrq     1.643 Ein-/Auswärts    6.551 IndirKollision     un  3.1  7.6
  zöä.ß bfgvx     18.561 benachbart       2.359 Shift-benachbart  sum 50.9 49.1
                 10.5  8.8 18.7 12.9 --.- --.- 15.3 14.4 10.5  8.9 Sh  1.8  1.0



AND THEN WE HAVE.... QWERTY

Qwerty           640.636 Gesamtaufwand  346.732 Lageaufwand        links rechts
                   6.804 Kollisionen      6.299 Shift-Kollisionen  ob 28.0 20.2
  qwert yuiopü    52.755 Handwechsel     41.474 Shift-Handwechsel  mi 22.1  9.5
  asdfg hjkläö     1.080 Ein-/Auswärts   11.226 IndirKollision     un  6.8 13.3
  zxcvb nm,.ß     21.628 benachbart      12.586 Shift-benachbart  sum 56.9 43.1
                  9.1  8.4 18.5 20.9 --.- --.- 18.4  8.9 12.1  3.6 Sh  1.1  1.7

Of course (by definition) the layout that the algorithm has calculated scores best, according to that same algorithm ;-)   But still, let's look at some figures.

DVORAK-GROUP
DvorMax beats Dvorak on nearly everything: much lower same finger (=Kollisionen), more alternation, less neighbouring fingers, smaller distance, better hand balance (49/51 versus 45/55), less load on pinkies. Hence, the overall score is much better. Dvorak has more inwards movements though. Klausler is nothing special. If you are after a Dvorak like layout, chose Dvorak for ubiquity, calculated your own

Standard ADNW (Aus der Neo Welt, made for 50% English, 50% German) scores worse than Dvormak (made for 100% English) on a 100% English corpus. Duh. If you use various langauges, like the Dvorak approach and don't want to calculate your own, ADNW is a very good choice. Also, it is in standard Linux distributions. 

COLEMAK GROUP
Colemak scores about the same as Dvorak (a little bit worse, actually), but the score is made up differently. Same finger is much better in Colemak than in Dvorak. I was surprised to see that it has less inward movement than Dvorak has. Inward rolls are one of the goals of Colemak, isn't it? Next there are some factors that ADNW sees differently than Colemak does. These are Handwechsel (=alternation) and Benachbart (=neigbouring fingers = rolls). ADNW gives penalties for the low alternation and high rate of rolls that Colemak has.  Capewell and Workman are nothing to write home about.  If you want it Colemaky, use Colemak :-) 

MIXED APPROACH GROUP
The MTGAP layout aims for both rolls and alternation. And it shows in the figures! In a sense it's in between Dvorak and Colemak. A very nice layout, really.

QWERTY
Qwerty is, as expected, bad.


That's all for now folks!

Last edited by pieter (19-Nov-2014 21:47:20)
Offline
  • 0
  • Reputation: 0
  • Registered: 15-Feb-2016
  • Posts: 12

This one has no name:


qcdOfuI.png

He optimized for more hand alternation and inward rolls with minimal outward rolls. 

OneProduct (my own. Need a name for it...)
A layout algorithmically generated based on heuristics. I used the same "simulated annealing" approach that Carpalx used, but with different heuristics. I then hand tweaked the results a bit so that it was a bit more "comfortable" as the heuristics don't account for human comfort as much as they could. I also made a few changes where the layout became more similar to some other layouts (for familiarity) where there was a very minimal decrease in heuristic viability. A few things in particular/of note:
1. Compared to Colemak, my current layout, the following are in the same places: GJEIRSTZVB. That's 10 out of 30.
2. ZCV remain on the bottom row near the left but because it is statistically good! I didn't have to force them to be there.
3. Compared to MTGAP, our algorithms came to the same conclusions a lot. In particular:
3.1 EO? in a column
3.2 A and U on the same finger
3.3 Punctuation on the same finger as A
3.4 A as the index finger key (this is important because the index fingers control 2 columns, and as such cause a lot of same finger clashes if not planned well)
3.5 RST arranged as an inward roll (colemak is the same)
3.6 IEA arranged as an inward roll
3.8 PNZ in a column

It appears to have been abandoned.

From: https://geekhack.org/index.php?topic=67604.0

Offline
  • 0