Serial Hill Climber in GoLang

Simple Binary Hill Climber in GoLang

I've been playing with Go recently and really enjoying it. But one of my problems was that I couldn't think of what to do with all that great concurrent power. Web applications and servers are of course one of the things I first thought of, but that's basically the Hello World of Go.

So instead, I looked through my books and spotted my copy of Josh Bongard's How the Body Shapes the Way We Think that I picked up when I took his course on evolutionary robotics. Our first assignment in that class was a serial hill climber. It was a super simple program written in python, a couple for loops and an object. The point of which was to mutate a genome string into the perfect fitness.

Generations? Population Sizes? Sounds perfect for some concurrency play! So I built an entire Evolutionary Computation Kit a small program to get a grip on how channels in Go work, and on wetting my appetite for Evolutionary methods.

Without further adieu, here's the entire Go program:

It's pretty simple. Here's a quick run down:

We create a simple struct to keep track of each member of the population, they have a genome as well as a fitness (based on said genome), and they also store a reference to a channel. If you're unfamiliar with Go, you can read up on them here or just take my word that they're like pipes in a Linux/Unix system.

To create a child we initialize it's genome to a random binary sequence of 1's and 0's. Simplistic yes, but this is a toy problem. A cool feature of the way we initialize our Child is that we use go routines to initialize each Child's genome concurrently; when we do this we also collect the initialized parts and feed them back into our calculation of the fitness as it happens. There's no real reason to do this in such a small example, but it's quite fun nonetheless.

func (c* Child) init(resChan * chan int) {
	c.resultChan = *resChan
	c.genome = make([]int, genomeSize)
	initilizeChannel := make(chan int, genomeSize)
	for i := 0; i < genomeSize; i++ {
		i := i //overshadow local 
		go func(){
			c.genome[i] = rand.Intn(2)
			initilizeChannel <- c.genome[i]
		}()	
	}
	//Compute initial fitness from initalization
	c.fitness = 0
	for i := 0; i < genomeSize; i++ {
		c.fitness += <- initilizeChannel
	}
}

The Mutate function represents our variation in the population, for this example we're merely flipping a coin to decide whether or not we should allow a single portion of the genomic string to change. One again, we fan out for the calculation and send the fitness back to the owning struct in parts. The most important part of the Mutate function though, is that it uses the internal resultChan channel variable in the Child to send the fitness result to whoever is listening.

func (c* Child) Mutate() {
	doneChannel := make(chan int, genomeSize)
	for i := 0; i < genomeSize; i++ {
		i := i
		go func(){
			if rand.Intn(2) == 1 {
				c.genome[i] = rand.Intn(2)
			}
			doneChannel <- c.genome[i]
		}()
	}
	c.fitness = 0
	for i := 0; i < genomeSize; i++ {
		c.fitness += <- doneChannel
	}
	//Send fitness out
	c.resultChan <- c.fitness
}

Where does resultChan go? To the channel we opened in the main function called resultsChan (name makes sense? imagine that). This channel is responsible for figuring out which of the population is the best candidate for reproduction to help shape the next generation of candidates. At the end each generation we mutate the best Child and replace a small subset of the population with them.

It's a fun toy example, and this is by no means a Go tutorial, but I hope that this helps inspire some ideas for people to start using Go as a language to express inheritly concurrent operations such as genetic algorithms and other evolutionary methods.