⌨️ Keyboard Configuration using Genetic Algorithm

Evolving an ergonomic keyboard layout that reduces typing effort by 49.4% (English) and 49.6% (Java) compared to QWERTY — powered by the Carpalx biomechanical model and real corpus data.

---

📌 Overview

The standard QWERTY keyboard was designed in 1873 for mechanical typewriters — not for human hands. It forces excessive finger travel, overloads weak pinky and ring fingers, and contributes to Repetitive Strain Injuries (RSI) such as Carpal Tunnel Syndrome and Tendonitis.

This project uses a Genetic Algorithm (GA) to discover a keyboard layout that minimizes biomechanical typing effort. The fitness function is grounded in the authentic Carpalx effort model — calibrated using real constants from published research — and trained on real corpus data for two use cases: English prose and Java code.

---

📊 Results

English Prose (books_short.txt — 2,373,664 characters)

Metric	QWERTY	GA Optimized
Carpalx effort score	3,521.37	1,781.59
Effort reduction	—	49.4%
Bigrams extracted	—	806
Trigrams used	—	6,978
Time to converge	—	~1,071s (1000 gen)

Optimized Layout — English

TOP : Q   J   G   F   V   ;   '   K   X   Z
HOME: B   C   L   N   M   I   A   D   Y   .
BOT : P   W   T   S   R   O   E   H   U   ,

Vowels I and A on the home row alongside high-frequency consonants L, N, M — with common letters T, S, R, O, E, H accessible on the bottom row.

---

Java Code (java.txt — 1,467,254 characters)

Metric	QWERTY	GA Optimized
Carpalx effort score	3,745.99	1,886.61
Effort reduction	—	49.6%
Bigrams extracted	—	861
Trigrams used	—	6,861
Time to converge	—	~1,128s (1000 gen)

Optimized Layout — Java

TOP : '   Q   J   B   X   ,   K   M   W   Z
HOME: Y   G   I   N   P   U   A   L   ;   V
BOT : H   D   E   R   C   O   S   T   .   F

Vowels I, U, A all on the home row. The semicolon ; pulled to home row (terminates every Java statement). Common Java keywords use E, R, C, S, T — all on the bottom row.

---

🧬 How It Works

The keyboard layout problem has 30! ≈ 2.65 × 10³² possible arrangements — far too large for exhaustive search. A Genetic Algorithm solves this by simulating natural evolution:

Random population (200 layouts)
        ↓
Evaluate fitness (Carpalx effort formula)
        ↓
Tournament selection (best of 5)
        ↓
OX1 Crossover (combine two parents)
        ↓
Swap mutation (8% chance)
        ↓
Repeat for 1,000 generations
        ↓
Best layout found

GA Concepts Mapped

GA Concept	Keyboard Mapping
Chromosome	One complete 30-key layout (Python list)
Gene	A single character at a key position
Population	200 randomly generated layouts
Fitness	1 / Carpalx effort score
Selection	Tournament — best of 5 random candidates
Crossover	Order Crossover OX1 — always valid permutation
Mutation	Swap two random keys (8% probability)
Elitism	Top 10% survive unchanged each generation

---

📐 Fitness Function

The fitness function uses the authentic Carpalx biomechanical effort model:

Fitness = 1 ÷ Σ ( K1·e₁ + K2·e₂ + K3·e₃ ) × freq(trigram)

Parameter	Value	Meaning
KB	0.3555	Base effort weight
KP	0.6423	Penalty effort weight
KS	0.4268	Stroke path weight
K1 / K2 / K3	1.000 / 0.367 / 0.235	Trigram position decay
W_ROW	1.3088	Row penalty multiplier
W_FINGER	2.5948	Finger penalty multiplier

Four components contribute to total cost:

1. Trigram path cost — K1/K2/K3 weighted stroke effort per 3-char sequence
2. Same-finger penalty — +1.5× when consecutive keys use the same finger
3. Same-hand penalty — +0.8 when all 3 trigram keys land on one hand
4. Hand balance penalty — scaled up to 200 for fully one-sided layouts

---

📚 Dataset — Carpalx Corpus

Trained on 14 classic English novels (~2.97 MB) from the Carpalx project:

Alice's Adventures in Wonderland	Dracula	Great Expectations
The Picture of Dorian Gray	Huckleberry Finn	Jane Eyre
Moby-Dick	The Count of Monte Cristo	Pride and Prejudice
Sense and Sensibility	Sherlock Holmes	Tom Sawyer
Ulysses	Walden

After processing: 2,373,664 characters → 806 bigrams → 6,978 trigrams

Java corpus: 1,467,254 characters → 861 bigrams → 6,861 trigrams

---

🗂️ Project Structure

keyboard-config-ga/
├── keyboard_ga.py                            # Main optimizer (Carpalx model + real corpus)
├── corpus/
│   ├── books_short.txt                       # English corpus
│   └── java.txt                              # Java corpus
├── outputs/
│   ├── keyboard_for_books.txt                # Full output — English run
│   └── keyboard_for_java.txt                 # Full output — Java run
│   ├── keyboard_results_carpalx.txt          # Output — optimized layout + cost history
├── assignment/
│   ├── assignment_keyboard_ga_final.docx     # Assignment write-up
│   └── keyboard_ga_presentation.pptx         # 12-slide presentation
├── carpalx-master/                           # Carpalx configs, corpora, and tools
├── carpalx.zip                               # Carpalx archive
├── Application_of_a_genetic_algorithm_to_the_keyboard.pdf
├── LICENSE
├── .gitignore
└── README.md

.gitignore excludes carpalx.zip, carpalx-master/, and Application_of_a_genetic_algorithm_to_the_keyboard.pdf from tracking.

---

⚙️ Setup & Usage

1. Clone the repo

git clone https://github.com/bharath-akurathi/keyboard-config-ga.git
cd keyboard-config-ga

2. Check Python version

python --version

Python 3.8+ required. No third-party packages needed — only random, math, collections, time from the standard library.

3. Verify corpus files

results/corpus/books_short.txt   # English prose corpus
results/corpus/java.txt          # Java code corpus

4. Run the optimizer

# Optimize for English prose
python results/keyboard_ga.py

# To optimize for Java code, change CORPUS_PATH in the script:
# CORPUS_PATH = "results/corpus/java.txt"

# Run alternate simplified implementation
python results/generated_data_ga.py

5. Expected output

[Corpus] Reading results/corpus/books_short.txt … 2,373,664 chars → 806 bigrams, 6,978 trigrams

==============================================================
  Carpalx GA  |  pop=200  gen=1000
==============================================================
  Gen     0 | Cost:    2574.76 | Elapsed:   1.1s
  Gen   100 | Cost:    1854.29 | Elapsed:  98.9s
  Gen   200 | Cost:    1808.35 | Elapsed: 212.7s
  Gen   300 | Cost:    1796.89 | Elapsed: 332.2s
  Gen   400 | Cost:    1794.06 | Elapsed: 446.8s
  Gen   500 | Cost:    1781.59 | Elapsed: 559.9s
  ...
  Gen   999 | Cost:    1781.59 | Elapsed: 1071.0s

Optimized Layout:
TOP : Q   J   G   F   V   ;   '   K   X   Z
HOME: B   C   L   N   M   I   A   D   Y   .
BOT : P   W   T   S   R   O   E   H   U   ,

QWERTY: 3521.37  |  Optimized: 1781.59  |  Reduction: 49.4%

---

🔧 Configuration

All parameters are at the top of results/keyboard_ga.py:

CORPUS_PATH     = "results/corpus/books_short.txt"  # swap to java.txt for code layout
POPULATION_SIZE = 200                # more = richer gene pool, slower
GENERATIONS     = 1000               # more = better result
MUTATION_RATE   = 0.08               # 8% — swap two keys per child
ELITE_FRACTION  = 0.10               # top 10% survive unchanged
TOURNAMENT_SIZE = 5                  # candidates per selection round

---

📈 Cost Over Generations

English corpus

Gen    0  →  2574.76   (random start)
Gen  100  →  1854.29   (fast early improvement)
Gen  200  →  1808.35
Gen  300  →  1796.89
Gen  400  →  1794.06
Gen  500  →  1781.59   (converged — 49.4% below QWERTY)
Gen 1000  →  1781.59   (held steady)

Java corpus

Gen    0  →  2609.25   (random start)
Gen  100  →  1953.64   (fast early improvement)
Gen  200  →  1911.04
Gen  300  →  1889.10
Gen  400  →  1886.61   (converged — 49.6% below QWERTY)
Gen 1000  →  1886.61   (held steady)

Both runs converged around generation 400–500 and held steady — a healthy sign the GA found a strong local optimum.

---

🏥 Why This Matters — Medical Context

Typing-related conditions affect millions of keyboard users:

• Carpal Tunnel Syndrome — median nerve compression from inflamed tendons
• Tendonitis / Tenosynovitis — tendon inflammation from repetitive movement
• Osteoarthritis aggravation — keypress micro-impacts worsen joint cartilage
• Trigger Finger — narrowed tendon sheath causes finger to lock

A layout that reduces effort by ~49% means significantly fewer micro-strain events per day — across millions of keystrokes per year, this makes a meaningful ergonomic difference.

---

📖 References

1. Krzywinski, M. (2007). Carpalx — Keyboard Layout Optimizer. Genome Sciences Centre. → http://mkweb.bcgsc.ca/carpalx/

2. Eggers, J., De Brock, B., & Land, R. (2020). Application of a Genetic Algorithm to the Keyboard Layout Problem. University of Groningen. → https://www.researchgate.net/publication/338443852

3. Norvig, P. (2012). English Letter Frequency Counts: Mayzner Revisited. Google Research. → http://norvig.com/mayzner.html

4. Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley.

5. adumb (2023). Using AI to Create the Perfect Keyboard [YouTube]. → https://youtu.be/EOaPb9wrgDY

---

📄 License

This project (code, assignment, and presentation) is released under the MIT License — free to use, modify, and distribute with attribution.

The Carpalx corpus (books) consists of public domain texts from Project Gutenberg. The Carpalx configuration files are the work of Martin Krzywinski. A local copy may be present in this working tree, but .gitignore is configured to exclude Carpalx archive/folder paths from new tracking.

---

_{Built as part of a Machine Learning assignment on Bio-Inspired Computing}