Show HN: A Genetic Algorithm I wrote in JavaScript to evolve "Hello World"

The neat part about it was that you could put in your own problem and have it solve it. I took a minute and slapped together the Hello World! version of this problem. You can copy it into the text box and play with the parameters (and resulting graph output. The output isn't a pretty as the link and it runs slower, but still neat.

function salesman() { this.target = "hello world!";

    this.points = [ 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', '!', ' ' ];

    this.fitness = function(chromosome) {
        // higher fitness is better
        var f = 0; // start at 0; the best fitness
        for(var i=0, c=this.target.length; i<c ; i++) {
            f -= Math.abs(this.target.charCodeAt(i) - this.points[chromosome[i]].charCodeAt(0));
        }
        return Math.abs(f);
     };
    
    // the size of values that should be passed to fitness
    this.numberOfArgs = function() { return this.target.length; };
    
    // the max value needed for the arguments
    this.maxArg = function() { return this.points.length; };
    
    // convert the current chromosome value which can have a maxValue
    // into something fitness can use.
    this.getArg = function(value, maxValue) {
        return Math.round(value * (this.points.length - 1) / maxValue);
    };

    // Paint the solution onto bestimage
    this.paint = function(values) {
         var canvas = document.getElementById('bestimage');
        if (canvas.getContext){
            var w = canvas.width;
            var h = canvas.height;
            var canvasContext = canvas.getContext('2d');
            canvasContext.clearRect(0, 0, w, h);
            
            canvasContext.font = "italic 200 12px/2 Unknown Font, sans-serif";
            canvasContext.strokeStyle = "blue";
            for (var i = 0; i < values.length; ++i){
                canvasContext.strokeText(this.points[values[i]], i*10, 10);
            }
        }
    }

in bash...

to generate C code...

using temp files to pass data around.

They didn't get far, but did actually manage to get something to print after a few hours, though it never seemed to get closer to "hello world" after that point. We figured it was just that the GA was a bit overly proud of its accomplishment, and wasn't interested in improvement. This is with random strings inside a main() wrapper with stdio, and nothing else. I forget what their fitness function was, but I'm fairly certain it wasn't too simple, and I think that shot them in the foot near the end.

I still boggle at the attempt; I built a maze solver in Ruby that handled changing mazes by ant-trail-like solving (I forget the term...) that only took me a handful of hours. Then I spent another 10 to make modifying the maze as simple as possible, and I made a dozen or so wildly different mazes with moving walls, transporters, etc while doing my presentation to demonstrate that it worked. I think that may have won me more points than the functioning code; lots of people's barely worked, and nobody else had anything they could run in reasonable times during a presentation.

[EDIT: s/GE/GA/ ... not sure why I got that wrong...]

One day I hope to apply machine learning to the art of predicting real-life events. This is one small step towards that project, and I'd love to hear your feedback! - But it's 1AM now so I'll be offline for the next few hours. Here's hoping I wake up to a million upvotes! :)

I did something similar in javascript --> https://github.com/JohnIdol/WasDarwinWrong ... and C# --> https://github.com/JohnIdol/typingmonkey

I tried 3 different types of fitness functions: Per-Character error summation, Levenshtein Distance and Hamming Distance.

I found that Hamming Distance produced the quickest convergences.

The code for my project is "fully OOP" so you can switch implementations for crossover, mutation and selection operators and also for the fitness functions.

Go ahead, fork me ;-)