correct skipgram computation

4 files changed, 54 insertions, 13 deletions

diff --git a/srchr/src/config.rs b/srchr/src/config.rs
index 2633a71..5bb35c5 100644
--- a/srchr/src/config.rs
+++ b/srchr/src/config.rs
@@ -1,19 +1,18 @@
 // With the below parameters and the supplied corpus searches seem to converge on a few layouts, with one of the better ones being
-
-// V L O M Q  P F U H ,
+// V L O M Z  P F U H ,
 // S R A T K  Y C E N I
-// Z X / D J  G W ' B .
+// X J / D Q  G W ' B .
 // Percent per key:
-//  0.92  3.96  7.42  2.57  0.10   1.62  2.10  2.77  5.94  1.78
+//  0.92  3.96  7.42  2.57  0.06   1.62  2.10  2.77  5.94  1.78
 //  6.03  5.56  7.74  8.60  0.77   1.99  2.35 11.90  6.62  6.68
-//  0.06  0.15  0.14  4.25  0.15   1.93  2.27  1.00  1.50  1.14
+//  0.15  0.15  0.14  4.25  0.10   1.93  2.27  1.00  1.50  1.14
 //
-// Finger usage:  7.02%,  9.66%, 15.30%, 16.44%, 12.25%, 15.67%, 14.06%,  9.61%
+// Finger usage:  7.10%,  9.66%, 15.30%, 16.36%, 12.25%, 15.67%, 14.06%,  9.61%
 // Hand usage: 48.42% vs 51.58%
-// SFB distribution :  0.001%,  0.087%,  0.089%,  0.047%,  0.104%,  0.197%,  0.044%,  0.024%
-// DSFB distribution:  0.147%,  0.431%,  0.615%,  0.999%,  0.208%,  0.997%,  1.102%,  0.408%
-// Total sfb: 0.594% (49972)
-// Total dsfb: 4.906% (412768)
+// SFB distribution :  0.002%,  0.088%,  0.089%,  0.047%,  0.104%,  0.197%,  0.044%,  0.024%
+// DSFB distribution:  0.166%,  0.402%,  0.614%,  1.302%,  0.761%,  0.985%,  1.101%,  0.405%
+// Total sfb: 0.595% (50086)
+// Total dsfb: 5.736% (482638)
 
 // -----------------------------------------------------------------------------
 //                                    Tweak
@@ -34,7 +33,8 @@ pub const MAX_NUM_TRANSPOSITIONS: usize = 12;
 
 // Fitness parameters
 pub const DESIRED_INDEX_USAGE_PERCENT: f32 = 0.12;
-pub const STANDARD_DSFB_WEIGHT: f32 = 1.0 / 32.0;
+pub const STANDARD_DSFB_WEIGHT: f32 = 1.0 / 8.0;
+pub const INDEX_DSFB_WEIGHT: f32 = 1.0 / 8.0;
 pub fn index_usage_fitness(total_count: u32, index_count: u32) -> f32 {
     let index_perc = index_count as f32 / total_count as f32;
     if index_perc < DESIRED_INDEX_USAGE_PERCENT {
diff --git a/srchr/src/corpus.rs b/srchr/src/corpus.rs
index 4428bc6..7216b1a 100644
--- a/srchr/src/corpus.rs
+++ b/srchr/src/corpus.rs
@@ -68,7 +68,7 @@ impl Corpus {
             if let Some(c) = current_char {
                 if let Some(lc) = last_char {
                     if let Some(llc) = last_last_char {
-                        if llc != c {
+                        if llc != lc && lc != c && llc != c {
                             skipgram_count[pair_to_index(llc, c)] += 1;
                             skipgram_count[pair_to_index(c, llc)] += 1;
                         }
diff --git a/srchr/src/evaluation.rs b/srchr/src/evaluation.rs
index a855d89..1c40b44 100644
--- a/srchr/src/evaluation.rs
+++ b/srchr/src/evaluation.rs
@@ -51,6 +51,23 @@ pub fn prelayout_fitness(corpus: &Corpus, layout: &Prelayout) -> f32 {
             + corpus.get_bigram_perc(k3, k4)
             + corpus.get_bigram_perc(k3, k5)
             + corpus.get_bigram_perc(k3, k6);
+
+        score += (corpus.get_skipgram_perc(k1, k2)
+            + corpus.get_skipgram_perc(k1, k3)
+            + corpus.get_skipgram_perc(k2, k3)
+            + corpus.get_skipgram_perc(k4, k5)
+            + corpus.get_skipgram_perc(k4, k6)
+            + corpus.get_skipgram_perc(k5, k6)
+            + corpus.get_skipgram_perc(k1, k4)
+            + corpus.get_skipgram_perc(k1, k5)
+            + corpus.get_skipgram_perc(k1, k6)
+            + corpus.get_skipgram_perc(k2, k4)
+            + corpus.get_skipgram_perc(k2, k5)
+            + corpus.get_skipgram_perc(k2, k6)
+            + corpus.get_skipgram_perc(k3, k4)
+            + corpus.get_skipgram_perc(k3, k5)
+            + corpus.get_skipgram_perc(k3, k6))
+            * INDEX_DSFB_WEIGHT;
     }
 
     score
@@ -105,6 +122,18 @@ impl Evaluation {
                     + corpus.get_bigram_perc(k3, k4)
                     + corpus.get_bigram_perc(k3, k5)
                     + corpus.get_bigram_perc(k3, k6);
+                dsfb[i] += corpus.get_skipgram_perc(k4, k5)
+                    + corpus.get_skipgram_perc(k4, k6)
+                    + corpus.get_skipgram_perc(k5, k6)
+                    + corpus.get_skipgram_perc(k1, k4)
+                    + corpus.get_skipgram_perc(k1, k5)
+                    + corpus.get_skipgram_perc(k1, k6)
+                    + corpus.get_skipgram_perc(k2, k4)
+                    + corpus.get_skipgram_perc(k2, k5)
+                    + corpus.get_skipgram_perc(k2, k6)
+                    + corpus.get_skipgram_perc(k3, k4)
+                    + corpus.get_skipgram_perc(k3, k5)
+                    + corpus.get_skipgram_perc(k3, k6);
             } else if i == 4 {
                 let k4 = layout.get_key(5 + 10 * 0);
                 let k5 = layout.get_key(5 + 10 * 1);
@@ -121,6 +150,18 @@ impl Evaluation {
                     + corpus.get_bigram_perc(k3, k4)
                     + corpus.get_bigram_perc(k3, k5)
                     + corpus.get_bigram_perc(k3, k6);
+                dsfb[i] += corpus.get_skipgram_perc(k4, k5)
+                    + corpus.get_skipgram_perc(k4, k6)
+                    + corpus.get_skipgram_perc(k5, k6)
+                    + corpus.get_skipgram_perc(k1, k4)
+                    + corpus.get_skipgram_perc(k1, k5)
+                    + corpus.get_skipgram_perc(k1, k6)
+                    + corpus.get_skipgram_perc(k2, k4)
+                    + corpus.get_skipgram_perc(k2, k5)
+                    + corpus.get_skipgram_perc(k2, k6)
+                    + corpus.get_skipgram_perc(k3, k4)
+                    + corpus.get_skipgram_perc(k3, k5)
+                    + corpus.get_skipgram_perc(k3, k6);
             }
         }
 
diff --git a/srchr/src/layout.rs b/srchr/src/layout.rs
index 841c085..81cfd9b 100644
--- a/srchr/src/layout.rs
+++ b/srchr/src/layout.rs
@@ -128,8 +128,8 @@ impl Layout {
                     wcol.sort_by(|(_, l), (_, r)| r.cmp(l));
                     wcol.swap(0, 1);
                     if N == 6 {
-                        wcol.swap(3, 4);
                         wcol.swap(0, 2);
+                        wcol.swap(3, 4);
                         wcol.swap(3, 5);
                     }
                     (wcol.into_iter().map(|(k, _)| k as u8).collect(), weight)