Playing Wumpus World with supervised learning using AI machine learning in R and caret. http://primaryobjects.github.io/wumpus

wumpus-1.csv

wumpus.R

library(caret)

# function to set up random seeds
setSeeds <- function(method = "cv", numbers = 1, repeats = 1, tunes = NULL, seed = 1237) {
  #B is the number of resamples and integer vector of M (numbers + tune length if any)
  B <- if (method == "cv") numbers
  else if(method == "repeatedcv") numbers * repeats
  else NULL

  if(is.null(length)) {
    seeds <- NULL
  } else {
    set.seed(seed = seed)
    seeds <- vector(mode = "list", length = B)
    seeds <- lapply(seeds, function(x) sample.int(n = 1000000, size = numbers + ifelse(is.null(tunes), 0, tunes)))
    seeds[[length(seeds) + 1]] <- sample.int(n = 1000000, size = 1)
  }
  # return seeds
  seeds
}

seed <- 1002
number <- 4
method <- 'cv'
cvSeeds <- setSeeds(method = method, numbers = number, tunes = 3, seed = seed)

# Control list for model training.
myControl <- trainControl(method = method, number = number, seeds = cvSeeds)

# Split the data into a training/test set by 70% training/30% test.
data <- read.csv('wumpus-1.csv')
inTrain <- createDataPartition(y = data$move, p=0.7, list=FALSE)
training <- data[inTrain,]
test <- data[-inTrain,]

# SVM: Correct: 6/19 = 32%
fit <- train(move ~ ., data = training, method = 'svmRadial', trControl = myControl)
results <- predict(fit, newdata = test)
table(results, test$move)
data.frame(test, results, test$move == results)
print(paste0('Correct: ', length(which(test$move == results)), '/', nrow(test), ' = ', round(length(which(test$move == results)) / nrow(test) * 100), '%'))

# Logistic regression: Correct: 8/19 = 42%
fit <- train(move ~ ., data = training, method = 'multinom', trControl = myControl)
results <- predict(fit, newdata = test)
table(results, test$move)
data.frame(test, results, test$move == results)
print(paste0('Correct: ', length(which(test$move == results)), '/', nrow(test), ' = ', round(length(which(test$move == results)) / nrow(test) * 100), '%'))

# K-nearest-neighbors: Correct: 10/19 = 53%
fit <- train(move ~ ., data = training, method = 'knn', trControl = myControl)
results <- predict(fit, newdata = test)
table(results, test$move)
data.frame(test, results, test$move == results)
print(paste0('Correct: ', length(which(test$move == results)), '/', nrow(test), ' = ', round(length(which(test$move == results)) / nrow(test) * 100), '%'))

# Random Forest: Correct: 16/19 = 84%
fit <- train(move ~ ., data = training, method = 'rf', trControl = myControl)
results <- predict(fit, newdata = test)
table(results, test$move)
data.frame(test, results, test$move == results)
print(paste0('Correct: ', length(which(test$move == results)), '/', nrow(test), ' = ', round(length(which(test$move == results)) / nrow(test) * 100), '%'))

# Regularized Random Forest: Correct: 16/19 = 84%
fit <- train(move ~ ., data = training, method = 'RRF', trControl = myControl)
results <- predict(fit, newdata = test)
table(results, test$move)
data.frame(test, results, test$move == results)
print(paste0('Correct: ', length(which(test$move == results)), '/', nrow(test), ' = ', round(length(which(test$move == results)) / nrow(test) * 100), '%'))

# Finally, train a model on the entire data-set so we can play the game: Correct: 67/68 = 99%
fit <- train(move ~ ., data = data, method = 'rf')
results <- predict(fit, newdata = data)
table(results, data$move)
data.frame(data, results, data$move == results)
print(paste0('Correct: ', length(which(data$move == results)), '/', nrow(data), ' = ', round(length(which(data$move == results)) / nrow(data) * 100), '%'))

# Manual test.
custom <- data.frame(room_n='visited', room_e='visited', room_s='visited', room_w='unvisited', breeze='false', stench='false', glitter='true')
predict(fit, newdata=custom)
# ^^ west (unvisited room)

# Actual play of the game.
custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='wall', room_w='wall', breeze='false', stench='false', glitter='false')
predict(fit, newdata=custom)
custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='visited', room_w='wall', breeze='true', stench='false', glitter='false')
predict(fit, newdata=custom)
custom <- data.frame(room_n='visited', room_e='unvisited', room_s='wall', room_w='wall', breeze='false', stench='false', glitter='false')
predict(fit, newdata=custom)
custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='wall', room_w='visited', breeze='true', stench='false', glitter='false')
predict(fit, newdata=custom)

# Example run of actual game from starting position.
# > custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='wall', room_w='wall', breeze='false', stench='false', glitter='false')
# > predict(fit, newdata=custom)
# [1] north
# Levels: east north south west
# ^^^ This is correct behavior, moving north when no obstacles detected.

#   > custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='visited', room_w='wall', breeze='true', stench='false', glitter='false')
# > predict(fit, newdata=custom)
# [1] south
# Levels: east north south west
# ^^^ This is correct behavior, moving back to the south room since a breeze was detected.

#   > custom <- data.frame(room_n='visited', room_e='unvisited', room_s='wall', room_w='wall', breeze='false', stench='false', glitter='false')
# > predict(fit, newdata=custom)
# [1] east
# Levels: east north south west
# ^^^ This is correct behavior, moving east to a newly unvisited room since we've already been north.

#   > custom <- data.frame(room_n='unvisited', room_e='unvisited', room_s='wall', room_w='visited', breeze='true', stench='false', glitter='false')
# > predict(fit, newdata=custom)
# [1] west
# Levels: east north south west
# ^^^ This is expected behavior, moving back to the west since a breeze was detected.