ANN

A MLP is a class of feedforward artificial neural network (ANN). The term MLP is used ambiguously, sometimes loosely to any feedforward ANN, sometimes strictly to refer to networks composed of multiple layers of perceptrons.

Notes:
mlp uses by default hyperbolic tangent as transfer function.

How to deal with the number of epochs, the stopping criteria and the optimization algorithm.
nnfor runs neuralnet package for R under the hood, which details all the training options. We can pass those to mlp() as well. Defaults work well.

Multiple steps ahead prediction
The multiple step forecast in nonlinear model is intractable because we loose normality after the first step. 
However it can still be acheived from bootstrap, Monte Carlo or Naive method.
Since we do not observed the complete nonlinearity captured by the network and we hope to remain in regions of values that you have more or less seen in the past. Then bootstrap is the case.

Address seasonality
- dummy variables (deterministic) and 
- lags (stochastic)

Linearity/Non-linearity
mpl is inherently nonlinear but it can approximate linear behaviours quite well

What algorithm is being used for training:
rprop: stands for resilient backpropagation, is a learning heuristic for supervised learning in feedforward artificial neural networks. This is a first-order optimization algorithm.
no need a validation set for training purposes

How to forecast with mlp usings xreg args.
1. Model x variable, forecast and feed y model with x (fit/actual+forecast)
2. Use lags of regressors
Limitations: 1. add uncertainty while 2. assumes that connection holds for y and x-lagged.