1. Load the occupancy train and test datasets in R:

# Load the occupancy data 
occupancy_train <-read.csv("C:/occupation_detection/datatraining.txt",stringsAsFactors = T)
occupancy_test <- read.csv("C:/occupation_detection/datatest.txt",stringsAsFactors = T)

# Define input (x) and output (y) variables
x = c("Temperature", "Humidity", "Light", "CO2", "HumidityRatio")
y = "Occupancy"

3. Based on the requirement by MXNet, convert the train and test datasets to a matrix and ensure that the class of the outcome variable is numeric (instead of factor as in the case of H2O):

# convert the train data into matrix
occupancy_train.x <- data.matrix(occupancy_train[,x])
occupancy_train.y <- occupancy_train$Occupancy

# convert the test data into matrix
occupancy_test.x <- data.matrix(occupancy_test[,x])
occupancy_test.y <- occupancy_test$Occupancy

4. Now, let's configure a neural network manually. First, configure a symbolic variable with a specific name. Then configure a symbolic fully connected network with five neurons in a single hidden layer followed with the softmax activation function with logit loss (or cross entropy loss). One can also create additional (fully connected) hidden layers with different activation functions.

# Configure Neural Network structure 
smb.data <- mx.symbol.Variable("data")
smb.fc <- mx.symbol.FullyConnected(smb.data, num_hidden=5) 
smb.soft <- mx.symbol.SoftmaxOutput(smb.fc)

5. Once the neural network is configured, let's create (or train) the (Feedforward) neural network model using the mx.model.FeedForward.create function. The model is fine-tuned for parameters such as the number of iterations or epochs (100), the metric for evaluation (classification accuracy), the size of each iteration or epoch (100 observations), and the learning rate (0.01):

# Train the network
model.nn <- mx.model.FeedForward.create(symbol = smb.soft,
                                        X = occupancy_train.x,
                                        y = occupancy_train.y,
                                        ctx = mx.cpu(),
                                        num.round = 100,
                                        eval.metric = mx.metric.accuracy,
                                        array.batch.size = 100,
                                        learning.rate = 0.01)