Kotlin Linear Regression with Gradient Descent

linear_regression_gradient_descent.kts

import java.net.URL

fun main(args: Array<String>) {

    // Fit points to a line conforming to y = mx + b

    class Point(val x: Double, val y: Double)

    val points = URL("https://tinyurl.com/yaxgfjzt")
            .readText().split(Regex("\\r?\\n"))
            .asSequence()
            .filter { it.isNotEmpty() }
            .map { it.split(",") }
            .map { Point(it[0].toDouble(), it[1].toDouble()) }
            .toList()

    // Points that exactly fit y = 2x + 100 
    // You'll want epochs = 1000000 and learningRate = .00001
    // val points = (1..100).map { Point(it, (it * 2) + 100) }

    val n = points.count().toDouble()

    // partial derivative with respect to M
    fun dM(expectedYs: List<Double>) =
            (-2.0 / n) * points.mapIndexed { i, p -> p.x * (p.y - expectedYs[i]) }.sum()

    // partial derivative with respect to B
    fun dB(expectedYs: List<Double>) =
            (-2.0 / n) * points.mapIndexed { i, p -> p.y - expectedYs[i] }.sum()

    val learningRate = .000001

    var m = 0.0
    var b = 0.0

    val epochs = 1000 // epochs is a fancy name for # of iterations

    (0..epochs).forEach { _ ->
        val yPredictions = points.map { (m * it.x) + b }

        val dM = dM(yPredictions)
        val dB = dB(yPredictions)

        m -= learningRate * dM
        b -= learningRate * dB
    }

    println("f(x) = ${m}x + $b")
    // RESULT:
    // f(x) = 1.6633912038847907x + 0.003908633274244349

    // View Graph on Desmos
    // https://www.desmos.com/calculator/ntomwigo6k
}