Status: This project is abandoned. I'm keeping it online because it may still prove to be useful for someone, but I will not address Issues or merge Pull requests.

CGP for Go

This is a library implementing Cartesian Genetic Programming in the Go programming language.

Installation

go get github.com/DataWraith/cgp

Usage

Here's an example of using CGP for symbolic regression. We are trying to approximate the function f(x) = x³ - 2x + 10.

import "fmt"
import "github.com/DataWraith/cgp"
import "math"
import "math/rand"

// First, we set up our parameters:
options := cgp.CGPOptions{
        PopSize:      100,  // The population size. One parent, 99 children
        NumGenes:     30,   // The maximum number of functions in the genome
        MutationRate: 0.05, // The mutation rate
        NumInputs:    1,    // The number of input values
        NumOutputs:   1,    // The number of output values
        MaxArity:     2,    // The maximum arity of the functions in the FunctionList
}

// The function list specifies the functions that are used in the genetic
// program. The first input is always the constant evolved for the function.
options.FunctionList = []cgp.CGPFunction{
        func(input []float64) float64 { return input[0] },            // Constant
        func(input []float64) float64 { return input[1] },            // Pass through A
        func(input []float64) float64 { return input[2] },            // Pass through B
        func(input []float64) float64 { return input[1] + input[2] }, // Addition
        func(input []float64) float64 { return input[1] - input[2] }, // Subtraction
        func(input []float64) float64 { return input[1] * input[2] }, // Multiplication

        // Division
        func(input []float64) float64 {
                if input[2] == 0 {
                        return 0
                }
                return input[1] / input[2]
        },
}

// This function is used to generate constants for use by the functions in
// the FunctionList. In this case we are using the provided instance of
// rand.Rand to generate a random constant between 0 and 100.
options.RandConst = func(rand *rand.Rand) float64 {
        return float64(rand.Intn(101))
}

// Prepare the testcases.
var testCases = []struct {
        in  float64
        out float64
}{
        {0, 10},
        {0.5, 9.125},
        {1, 9},
        {10, 990},
        {-5, -105},
        {17, 4889},
        {3.14, 34.679144},
}

// The evaluator uses the test cases to grade an individual by setting the
// fitness to the sum of squared errors. The lower the fitness the better the
// individual. Note that the input to the individual is a slice of float64.
options.Evaluator = func(ind cgp.Individual) float64 {
       fitness := 0.0
       for _, tc := range testCases {
                input := []float64{tc.in}
                output := ind.Run(input)
                fitness += math.Pow(tc.out-output[0], 2)
        }
        return fitness
}

// Initialize CGP
gp := cgp.New(options)

// Population[0] is the parent, which is the most fit individual. We
// loop until we've found a perfect solution (fitness 0)
for gp.Population[0].Fitness > 0 {
        gp.RunGeneration()
        fmt.Println(gp.Population[0].Fitness)
}