pragmatics.pjs

/*
testing the nested-query pragmatics model in js
*/

//var util = require("./probabilistic/util")
//util.openModule(util)
//var pr = require("./probabilistic/index")
//util.openModule(pr)


/////////////////////////////
////there is a useful pattern. if we have a marginalized fn:
//var foo = marginalize(prob(function(x) { ...}))
////and we condition on it returning a particular value:
//condition(foo(x) == val)
////we should avoid sampling the foo ERP... instead use isConditioned:
//var foo = marginalizeConditioned(prob(function(x) { ...}))
//condition(foo(x, val) == val)
////this avoids setting and resampling the conditioned ERP. (does not work with rejection sampling!)
/////////////////////////////

var meaning = function(utterance)
{
    if(utterance == "foo"){return function(w){return w}} //"foo" is true only in true worlds.
    return function(w){return true} //"bar" is always true
}

var listener = function(utterance) {
    var world = flip(); //world prior
    //speaker(world, utterance)
    condition(speaker.conditionTo(utterance)(world) == utterance)
    return world
}


var speaker = marginalize(
                       function(world) {
                          var utterance = flip() ? "foo" : "bar"
                          //litlistener(utterance, world)
                          condition(litlistener.conditionTo(world)(utterance) == world)
                          return utterance
                          }, samplefn, samplesPerLevel)


var litlistener = marginalize(
                            function(utterance)
                              {
                              var world = flip(); //world prior
                              condition(meaning(utterance)(world)) //simple stand-in for truth-functional meaning..
                              return world
                              }, samplefn, samplesPerLevel)

/////////////////////////////


//means=[]
//for(i=0;i<10;i++)
//{
//    var hist = distrib( ,
//                       traceMH,
//                       100)
//    console.log(hist[0])
//    means[i] = hist[0]
//}

var comp = function() {return listener("bar")}

var samplefn = traceMH //bunchaRejectionSample

var samplesPerLevel = 500

function variance(values)
{
    sq = []
    m = mean(values)
    for(i=0; i<values.length; i++)
    {
        sq.push( (values[i] - m)*(values[i] - m) )
    }
    
    return Math.sqrt(mean(sq))
}

function estimate() {
    litlistener.clearCache()
    speaker.clearCache()
    return expectation(comp, samplefn, samplesPerLevel)
}

var runs = repeat(100, estimate)

console.log( "mean: " + mean(runs) + " variance: " + variance(runs) )