Update for nnfor: reuse and retrain models

By | December 11, 2017

A feature that was missing in the nnfor package was the ability to reuse and retrain mlp and elm model fits. This is now possible with the new arguments model and retrain.

As an example, let us use the AirPassengers time series, with three different sample sizes and re-use and re-train the same model in various combinations.

# Get some data
y <- AirPassengers
y1 <- window(y,end=c(1958,12))
y2 <- window(y,end=c(1959,12))
y3 <- window(y,end=c(1960,12))

# Fit NN 
fit <- list()
fit[[1]] <- mlp(y1)
fit[[2]] <- mlp(y2,model=fit[[1]])
fit[[3]] <- mlp(y2,model=fit[[1]],retrain=TRUE)
fit[[4]] <- mlp(y3,model=fit[[1]])
fit[[5]] <- mlp(y3,model=fit[[3]])
fit[[6]] <- mlp(y3,model=fit[[1]],retrain=TRUE)
fit[[7]] <- mlp(y3)

# Get MSE and number of lags
mse <- unlist(lapply(fit,function(x){x$MSE}))
lags <- unlist(lapply(fit,function(x){length(x$lags)}))
Model Series Sample Training Sample Retrain MSE Lags
fit[[1]] up to 1958 up to 1958 X 6.73 9
fit[[2]] up to 1958 up to 1958 61.25 9
fit[[3]] up to 1959 up to 1959 X 6.68 9
fit[[4]] up to 1960 up to 1958 541.13 9
fit[[5]] up to 1960 up to 1959 260.22 9
fit[[6]] up to 1960 up to 1960 X 12.65 9
fit[[7]] up to 1960 up to 1960 New fit 7.95 10

As you can see, with different in-sample data and no retraining the in-sample MSE deteriorates. Using the “up to 1958” fit on the “up to 1959” and “up to 1960” samples returns MSE of 61.25 and 541.13 respectively. If we refit the network (keeping the specification fixed) the error reduces to 6.68 and 12.65 respectively. Building a new model on the “up to 1960” data finds a different lag structure (increasing the used lags from 9 to 10), resulting in an MSE of 7.95.

The same arguments can be used with elm.

Leave a Reply

Your email address will not be published. Required fields are marked *