webppl/src/dists/logitNormal.js

Probabilistic programming for the web

'use strict';
var ad = require('../ad');
var base = require('./base');
var types = require('../types');
var gaussian = require('./gaussian');
var Gaussian = gaussian.Gaussian;
var LogitNormal = base.makeDistributionType({
    name: 'LogitNormal',
    desc: 'A distribution over ``(a,b)`` obtained by scaling and shifting a standard logit-normal.',
    params: [
        {
            name: 'mu',
            desc: 'location',
            type: types.unboundedReal
        },
        {
            name: 'sigma',
            desc: 'scale',
            type: types.positiveReal
        },
        {
            name: 'a',
            desc: 'lower bound',
            type: types.unboundedReal
        },
        {
            name: 'b',
            desc: 'upper bound (>a)',
            type: types.unboundedReal
        }
    ],
    wikipedia: 'Logit-normal_distribution',
    mixins: [base.continuousSupport],
    sample: function () {
        var a = ad.value(this.params.a);
        var b = ad.value(this.params.b);
        var mu = ad.value(this.params.mu);
        var sigma = ad.value(this.params.sigma);
        var x = gaussian.sample(mu, sigma);
        return ad.scalar.sigmoid(x) * (b - a) + a;
    },
    score: function (val) {
        var a = this.params.a;
        var b = this.params.b;
        var y = ad.scalar.div(ad.scalar.sub(val, a), ad.scalar.sub(b, a));
        var x = ad.scalar.log(ad.scalar.div(y, ad.scalar.sub(1, y)));
        var gaussScore = gaussian.score(this.params.mu, this.params.sigma, x);
        return ad.scalar.sub(gaussScore, ad.scalar.log(ad.scalar.mul(ad.scalar.mul(y, ad.scalar.sub(1, y)), ad.scalar.sub(b, a))));
    },
    base: function () {
        return new Gaussian({
            mu: 0,
            sigma: 1
        });
    },
    transform: function (x) {
        var a = this.params.a;
        var b = this.params.b;
        var mu = this.params.mu;
        var sigma = this.params.sigma;
        return ad.scalar.add(ad.scalar.mul(ad.scalar.sigmoid(ad.scalar.add(ad.scalar.mul(x, sigma), mu)), ad.scalar.sub(b, a)), a);
    },
    support: function () {
        return {
            lower: this.params.a,
            upper: this.params.b
        };
    }
});
module.exports = { LogitNormal: LogitNormal };