library(testthat)

monkeydat <- data.frame(Monkey=c("DAN", "ART", "CAU", "DOR", "VIO", "ARI"),
                       NumTrials=c(56689, 50985, 61142, 49608, 43041, 55407),
                       NumWordsLearned=c(308, 125, 112, 121, 81, 87),
                       NumNonwordsPresented=c(7832, 7832, 7832, 7832, 7832, 7832),
                       GeneralAccuracy=c(79.81, 73.41, 72.43, 73.15, 71.55, 71.14),
                       WordAccuracy=c(80.01, 74.83, 73.15, 79.26, 76.75, 75.38),
                       NonwordAccuracy=c(79.61, 72.00, 71.72, 67.06, 66.33, 66.90),
                       stringsAsFactors=F)

Resample <- function(x){
    x[sample.int(length(x))]
}

test_that("Resample", {
   x <- 1:10
   expect_that(length(Resample(x[x >  8])), equals(2))
   expect_that(length(Resample(x[x >  9])), equals(1))
   expect_that(length(Resample(x[x >  10])), equals(0))
})

RandomPick <- function(data, n) {
    if ( is.data.frame(data) ){
        data[sample(nrow(data), n, replace=T),]
    } else {
        data[sample(length(data), n, replace=T)]
    }
}

test_that("RandomPick", {
    n <- 10

    bigger <- 1:(3 * n)
    equal <- 1:n
    smaller <- c(1)
    empty <- c()
    df <- data.frame(a=c(1,2,3), b=c(4,5,6))

    expect_equal(length(RandomPick(bigger, n)), n)
    expect_equal(length(RandomPick(equal, n)), n)
    expect_equal(length(RandomPick(smaller, n)), n)
    expect_equal(length(RandomPick(empty, n)), 0)
    expect_equal(nrow(RandomPick(df, n)), n)
})


CreateBlock <- function(data){
    result <- data.frame(Stimulus=character(BLOCK_SIZE),
                         StimulusType=factor(character(BLOCK_SIZE),
                                             c("LearnedWord","NewWord","Nonword")),
                         stringsAsFactors=F)
    data <- data[,c("Stimulus","StimulusType")]
    stopifnot(is.character(data$Stimulus))

    newSample <- RandomPick(data[data$StimulusType == "NewWord",], 1)
    learnedSample <- RandomPick(data[data$StimulusType == "LearnedWord",], 25)
    if( nrow(newSample) > 0 ) {
        result[1:25,] <- newSample
    } else {
        result[1:25,] <- learnedSample
    }
    if( nrow(learnedSample) > 0 ){
        result[26:50,] <- learnedSample
    } else {
        result[26:50,] <- newSample
    }
   
    result[51:100,] <- RandomPick(data[data$StimulusType == "Nonword",], 50)

    result <- result[sample(nrow(result)),]
    rownames(result) <- 1:nrow(result)
    result
}

BLOCK_SIZE <- 100
test_that("CreateBlock", {
    stim <- as.character(1:300)
    data <- data.frame(Stimulus=stim, stringsAsFactors=F)
    data$StimulusType[1:100] <- "NewWord"
    data$StimulusType[101:200] <- "LearnedWord"
    data$StimulusType[201:300] <- "Nonword"

    block <- CreateBlock(data)
    trialNewWords <- block$Stimulus[ block$StimulusType == "NewWord" ]
    trialLearnedWords <- block$Stimulus[ block$StimulusType == "LearnedWord" ]
    trialNonwords <- block$Stimulus[ block$StimulusType == "Nonword" ]

    expect_that(nrow(block), equals(BLOCK_SIZE))
    expect_that(length(trialNewWords), equals(25))
    expect_that(length(unique(trialNewWords)), equals(1))
    expect_that(length(trialLearnedWords), equals(25))
    expect_that(length(trialNonwords), equals(50))
    # TODO: Decide what should be expected for empty vectors?
})

PresentTrials <- function(trialCount, present, data){
    trials <- data.frame()
    for ( curTrialNum in 1:trialCount ) {
        curBlock <- ( (curTrialNum - 1) %/% BLOCK_SIZE) + 1
        curTrialNumInBlock <- ( (curTrialNum - 1) %% BLOCK_SIZE ) + 1
        isNewBlock <- curTrialNumInBlock == 1

        if ( isNewBlock ){
            isNotFirstBlock <- curBlock > 1
            if ( isNotFirstBlock ){
                newWordTrials <- block[block$StimulusType == "NewWord",]
                if( length(unique(newWordTrials$Stimulus)) == 1 ){
                    newWordTrials <- newWordTrials[1,]
                    wordResponses <- newWordTrials[newWordTrials$Response == "Word",]
                    wordResponseRate <- nrow(wordResponses) / nrow(newWordTrials)
                    if ( wordResponseRate >= 0.8  ){
                        data$StimulusType[data$Stimulus == newWordTrials$Stimulus] <- "LearnedWord"
                    }
                }
            }
            block <- CreateBlock(data)
            block$Block <- curBlock
            block$TrialInBlock <- 1:nrow(block)
        }

        curTrial <- block[curTrialNumInBlock,]
        curStim <- curTrial$Stimulus
        curStimIsWord <- present(curStim, ifelse(curTrial$StimulusType == "Nonword", "Nonword", "Word"))

        block$Trial[curTrialNumInBlock] <- curTrialNum
        block$Response[curTrialNumInBlock] <- curStimIsWord

        trials <- rbind(trials, block[curTrialNumInBlock,])
    }
    rownames(trials) <- 1:trialCount
    result <- list(Trials=trials,Stimuli=data)
    result
}

test_that("PresentTrials", {
    stim <- as.character(1:300)
    data <- data.frame(Stimulus=stim, stringsAsFactors=F)
    data$StimulusType[1:100] <- "NewWord"
    data$StimulusType[101:200] <- "LearnedWord"
    data$StimulusType[201:300] <- "Nonword"
    count <- BLOCK_SIZE * 2.5

    result <- PresentTrials(count,
                            function(cue, outcome){
                                return(outcome)
                            },
                            data)
    trials <- result$Trials
    stimuli <- result$Stimuli
    expect_equal(nrow(trials), count)
    expect_equal(length(stimuli$StimulusType[stimuli$StimulusType == "LearnedWord"]), 102)
})

# The original colSums fails if matrix filters to 1 row and gets implicit castet to a vector
ColSums <- function(x) { 
    if(is.matrix(x)){
       return(colSums(x))
    } else {
        return(x)
    }
}

test_that("ColSums",{
    m2x2 <- matrix(1,2,2)
    colnames(m2x2) <- c("a","b")

    expect_equal(ColSums(m2x2), c(a=2,b=2))
    expect_equal(ColSums(m2x2[1,]), c(a=1,b=1))
})

MakeMonkey <- function(cueset, outcomeset=c("Word", "Nonword"),
                       alpha=sqrt(0.001), beta=sqrt(0.001), lambda=1.0){
    weights <- matrix(0, length(cueset), length(outcomeset))
    rownames(weights) <- cueset
    colnames(weights) <- outcomeset

    rate <- function(cue){
        cues <- unlist(strsplit(cue, "_"))
        cueWeights <- weights[cues, outcomeset]
        totalActivation <- ColSums(cueWeights)

        maxActivated <- names(totalActivation[totalActivation == max(totalActivation)])
        return(RandomPick(maxActivated,1))
    }
    
    learner <- function(stim, resp){
        cues <- unlist(strsplit(stim, "_"))

        cueWeights <- weights[cues, outcomeset]
        totalActivation <- ColSums(cueWeights)
       
        for (j in 1:length(cues)) {
            if (resp == "Nonword") {
                yesType <- "Nonword"
                noType <- "Word"
            } else if (resp == "Word") {
                yesType <- "Word"
                noType <- "Nonword"
            } else {
                stop("Unknown outcome", resp)
            }
            weights[cues[j],yesType] <<-  weights[cues[j],yesType] +
                alpha * beta * (lambda - totalActivation[yesType])
            weights[cues[j],noType] <<-  weights[cues[j],noType] +
                alpha * beta * (0 - totalActivation[noType])
        }
        return(rate(stim))
    }

    give_weights <- function(){
        return(weights)
    }

    list(
        GetWeights=give_weights,
        Learner=learner,
        Rate=rate
    )
}

test_that("MakeMonkey",{
    cueset <- c("a","b","c")

    monkey <- MakeMonkey(cueset)
    for (i in 1:1000){
        monkey$Learner("a_b","Word")
        monkey$Learner("b_c","Nonword")
    }
 
    print(monkey$GetWeights())
    expect_equal(monkey$Rate("a"), "Word")
    expect_equal(monkey$Rate("c"), "Nonword")
})

library(ndl)
dat <- read.table("dataDan.txt", header=T, stringsAsFactors=F)
dat$Cues <- orthoCoding(dat$String, grams=2)
dat$Frequency <- 1
dat$Stimulus <- dat$Cues
dat$StimulusType <- factor(ifelse(dat$Type == "word", "NewWord", "Nonword"), c("Nonword","NewWord","LearnedWord"))
dat$Outcomes <- factor(ifelse(dat$Type == "word", "Word", "Nonword"), c("Word","Nonword"))

dat <- dat[sample(1:nrow(dat)),]
cues <- unique(unlist(strsplit(dat$Cues, "_")))
outcomes <- levels(dat$Outcomes)
monkey <- MakeMonkey(cues, outcomes, alpha=sqrt(0.001), beta=sqrt(0.001), lambda=1.0)
monkeyName <- "DAN"
trialNum <- round(0.5 * as.numeric(monkeydat[monkeydat$Monkey == monkeyName,"NumTrials"]))
system.time(pres <- PresentTrials(
    trialNum
   ,monkey$Learner,dat))

# Analyse data:
learnedWordNum <- nrow(pres$Stimuli[pres$Stimuli$StimulusType == "LearnedWord",])
trialNum <- nrow(pres$Trials)
nonwordTrials <- pres$Trials[pres$Trials$StimulusType == "Nonword",]
wordTrials <- pres$Trials[pres$Trials$StimulusType != "Nonword",]
presNonwordNum <- length(unique(nonwordTrials$Stimulus))
presWordNum <- length(unique(wordTrials$Stimulus))
nonwordAcc <- nrow(nonwordTrials[nonwordTrials$Response == "Nonword",]) / nrow(nonwordTrials)
wordAcc <- nrow(nonwordTrials[wordTrials$Response == "Word",]) / nrow(wordTrials)
genAcc <- (nrow(nonwordTrials) * nonwordAcc + nrow(wordTrials) * wordAcc) / trialNum

monkeydat <- rbind(monkeydat,
                   c(paste(monkeyName, "Sim1"),
                     trialNum,
                     learnedWordNum,
                     presNonwordNum,
                     genAcc,
                     wordAcc,
                     nonwordAcc))

# TODO: Das in schöne Funktionen packen ;)

# TODO: Durchlaufen lassen für ca 50k Trials für verschiedene alpha, beta zwischen 0 und 1
# TODO: Plotten der verschiedenen Ergebnisse (num,acc) als heatmap 2d plot.
# TODO: Plotten der Unterschiede zu den Vorgabeaffen und optimale alpha beta für diese finden.

# what would be the theoretic success rates?
wRW <- monkey$GetWeights()
aRW <- estimateActivations(dat, wRW)$activationMatrix
aRW <- aRW[,order(colnames(aRW))]
dat$ChoiceRW <- apply(aRW, 1, FUN=function(v){
                   if(v["Nonword"] >= v["Word"]) {
                     return("Nonword")
                   } else {
                     return("Word")
                   }})
table(dat$Type, dat$ChoiceRW)


# EOF