validann {validann}

Validate Artificial Neural Networks.

Description

Compute metrics and statistics for predictive, replicative and/or structural validation of artificial neural networks (ANNs).

Usage

validann(...)

## S3 method for class 'ann'
validann(net, obs = NULL, sim = NULL, x = NULL,
  na.rm = TRUE, ...)

## S3 method for class 'nnet'
validann(net, obs = NULL, sim = NULL, x = NULL,
  na.rm = TRUE, ...)

## Default S3 method:
validann(obs, sim, wts = NULL, nodes = NULL,
  na.rm = TRUE, ...)

Arguments

net

an object of class ‘ann’ (as returned by function ann) or ‘nnet’ (as returned using nnet). This is a list object comprising information about the fitted ANN model, including values of weights, fitted target values, number of layers and numbers of nodes in each layer, for example.

obs, sim

vectors comprising observed (obs) and simulated (sim) examples of a single response variable. These vectors are used to compute model fit statistics. Optional if net is supplied (see ‘Details’).

x

matrix, data frame or vector of input data used for fitting net object. A vector is considered to comprise examples of a single input or predictor variable. While x is optional, sensitivity analyses useful for structural validation cannot be performed if it is not supplied.

na.rm

logical; should missing values (including NaN) be removed from calculations? Default = TRUE.

wts

vector of ANN weights used to compute input ‘relative importance’ measures if net object is not supplied. Must be supplied together with nodes in order to compute such metrics. See ‘Details’ for ordering of wts vector.

nodes

vector indicating the number of nodes in each layer of the ANN model. This vector should have 3 elements: nodes in input layer, nodes in hidden layer (can be 0), and nodes in output layer. If net object is not supplied, nodes must be supplied together with wts if any structural validation metrics are to be computed.

...

arguments to be passed to different validann methods, see specific formulations for details.

Details

To compute all possible validation metrics and statistics, net must be supplied and must be of class ‘ann’ (as returned by ann) or ‘nnet’ (as returned by nnet). However, a partial derivative (PaD) sensitivity analysis (useful for structural validation) will only be carried out if net is of class ‘ann’.

If obs and sim data are supplied in addition to net, validation metrics are computed based on these. Otherwise, metrics and statistics are computed based on obs and sim datasets derived from the net object (i.e. the data used to fit net and the fitted values). As such, both obs and sim must be supplied if validation is to be based either on data not used for training or on unprocessed training data (if training data were preprocessed). If either obs or sim is specified but the other isn't, both obs and sim will be derived from net if supplied (and a warning will be given). Similarly, this will occur if obs and sim are of different lengths.

If net is not supplied, both obs and sim are required. This may be necessary if validating an ANN model not built using either the nnet or ann functions. In this case, both wts and nodes are also required if any structural validation metrics are to be returned. If an ANN model has K input nodes, J hidden nodes and a single output O, with a bias node for both the hidden and output layers, the wts vector must be ordered as follows:

c(Wi1h1,Wi1h2,...Wi1hJ,Wi2h1,...Wi2hJ,...,WiKh1,...,WiKhJ,Wi0h1,...,Wi0hJ,
Wh1O,...,WhJO,Wh0O)

where Wikhj is the weight between the kth input and the jth hidden node and WhjO is the weight between the jth hidden node and the output. The bias weight on the jth hidden layer node is labelled Wi0hj while the bias weight on the output is labelled Wh0O. The wts vector assumes the network is fully connected; however, missing connections may be substituted by zero weights. Skip-layer connections are not allowed.

Value

list object of class ‘validann’ with components dependent on arguments passed to validann function:

metrics

a data frame consisting of metrics:

AME, PDIFF, MAE, ME, RMSE, R4MS4E, AIC, BIC, NSC, RAE, PEP, MARE, MdAPE, MRE, MSRE, RVE, RSqr, IoAd, CE, PI, MSLE, MSDE, IRMSE, VE, KGE, SSE and R.

See Dawson et al. (2007) for definitions.

obs_stats

a data frame consisting of summary statistics about the obs dataset including mean, minimum, maximum, variance, standard deviation, skewness and kurtosis.

sim_stats

a data frame consisting of summary statistics about the sim dataset including mean, minimum, maximum, variance, standard deviation, skewness and kurtosis.

residuals

a 1-column matrix of model residuals (sim - obs).

resid_stats

a data frame consisting of summary statistics about the model residuals including mean, minimum, maximum, variance, standard deviation, skewness and kurtosis.

ri

a data frame consisting of ‘relative importance’ values for each input. Only returned if net or wts and nodes are supplied.

If net is supplied, relative importance values computed using the following 4 methods are returned:

Garson's (Garson); connection weight (CW); Profile sensitivity analysis (Profile); and partial derivative sensitivity analysis (PaD).

In addition, if net is of class ‘ann’ (as returned by function ann) and the activation function used at the hidden layer (act_hid) is "tanh", relative importance values computed using the modified CW (MCW) are also returned. This method requires that the hidden layer activation function be symmetric about the origin.

If wts and nodes are supplied, only relative importance values computed using the Garson and CW methods are returned.

See Gevrey et al. (2003), Olden et al. (2004) and Kingston et al. (2006) for details of the relative importance methods.

y_hat

a matrix of dimension c(101, ncol(x) * 6) of model response values indicating the local sensitivity of the model to each input in x. Only returned if net and x are supplied.

The response values returned in y_hat are calculated using the ‘Profile’ sensitivity analysis method described in Gevrey et al. (2003). Using this method, the local sensitivity of each input in x is considered successively. For each input x[,i], 5 synthetic data sets are generated where inputs x[,-i] are successively fixed at their minimum, 1st quartile, median, 3rd quartile and maximum values (as calculated from x), while input x[,i] is varied between its minimum and maximum value, increasing in increments of 1% (giving 101 synthetic values of x[,i] for each of the 5 sets of fixed x[,-i]). These data are input into net and model response values corresponding to the 5 summary statistics are computed. These 5 sets of response values, together with a set of computed median responses, are returned as y_hat[,(i - 1) * 6 + 1:6]. This process is repeated for each input variable in x. See Gevrey et al. (2003) for further details.

as

a matrix of dimension dim(x) of ‘absolute sensitivity’ values for each input in x given the model output values (i.e. sim). Only returned if net and x are supplied and net is of class ‘ann’.

The values in as are calculated according to the partial derivative (PaD) sensitivity analysis method described in Gevrey et al. (2003), which involves computing the first-order partial derivatives of the ANN output with respect to each input. net must be of class ‘ann’ in order to access partial derivatives of the hidden layer nodes as returned by ann.

rs

a matrix of dimension dim(x) of ‘relative sensitivity’ values for each input in x given the model output values (i.e. sim). Only returned if net and x are supplied and net is of class ‘ann’.

To compute the values in rs, the as values are normalised by multiplying by x[,i]/sim as described in Mount et al. (2013). As for as, net must be of class ‘ann’ in order to access partial derivatives of the hidden layer nodes as returned by ann.

Methods (by class)

• ann: Compute validation metrics when net is of class ‘ann’.

• nnet: Compute validation metrics when net is of class ‘nnet’.

• default: Useful for predictive validation only or when ANN model has not been developed using either ann or nnet. Limited structural validation metrics may be computed and only if wts and nodes are supplied.

References

Dawson, C.W., Abrahart, R.J., See, L.M., 2007. HydroTest: A web-based toolbox of evaluation metrics for the standardised assessment of hydrological forecasts. Environmental Modelling & Software, 22(7), 1034-1052. http://dx.doi.org/10.1016/j.envsoft.2006.06.008.

Olden, J.D., Joy, M.K., Death, R.G., 2004. An accurate comparison of methods for quantifying variable importance in artificial neural networks using simulated data. Ecological Modelling 178, 389-397. http://dx.doi.org/10.1016/j.ecolmodel.2004.03.013.

Gevrey, M., Dimopoulos, I., Lek, S., 2003. Review and comparison of methods to study the contribution of variables in artificial neural network models. Ecological Modelling 160, 249-264. http://dx.doi.org/10.1016/S0304-3800(02)00257-0.

Kingston, G.B., Maier, H.R., Lambert, M.F., 2006. Forecasting cyanobacteria with Bayesian and deterministic artificial neural networks, in: IJCNN '06. International Joint Conference on Neural Networks, 2006., IEEE. pp. 4870-4877. http://dx.doi.org/10.1109/ijcnn.2006.247166.

Mount, N.J., Dawson, C.W., Abrahart, R.J., 2013. Legitimising data-driven models: exemplification of a new data-driven mechanistic modelling framework. Hydrology and Earth System Sciences 17, 2827-2843. http://dx.doi.org/10.5194/hess-17-2827-2013.

See Also

ann, plot.validann, predict.ann

Examples

# get validation results for 1-hidden node `ann' model fitted to ar9 data
# based on training data.
# ---
data("ar9")
samp <- sample(1:1000, 200)
y <- ar9[samp, ncol(ar9)]
x <- ar9[samp, -ncol(ar9)]
x <- x[, c(1,4,9)]

fit <- ann(x, y, size = 1, act_hid = "tanh", act_out = "linear", rang = 0.1)
results <- validann(fit, x = x)

# get validation results for above model based on a new sample of ar9 data.
# ---
samp <- sample(1:1000, 200)
y <- ar9[samp, ncol(ar9)]
x <- ar9[samp, -ncol(ar9)]
x <- x[, c(1,4,9)]

obs <- y
sim <- predict(fit, newdata = x)
results <- validann(fit, obs = obs, sim = sim, x = x)

# get validation results for `obs' and `sim' data without ANN model.
# In this example `sim' is generated using a linear model. No structural
# validation of the model is possible, but `wts' are provided to compute the
# number of model parameters needed for the calculation of certain
# goodness-of-fit metrics.
# ---
samp <- sample(1:1000, 200)
y <- ar9[samp, ncol(ar9)]
x <- ar9[samp, -ncol(ar9)]
x <- as.matrix(x[, c(1,4,9)])
lmfit <- lm.fit(x, y)
sim <- lmfit$fitted.values
obs <- y
results <- validann(obs = obs, sim = sim, wts = lmfit$coefficients)

# validann would be called in the same way if the ANN model used to generate
# `sim' was not available or was not of class `ann' or `nnet'. Ideally in
# this case, however, both `wts' and `nodes' should be supplied such that
# some structural validation metrics may be computed.
# ---
obs <- c(0.257, -0.891, -1.710, -0.575, -1.668, 0.851, -0.350, -1.313,
         -2.469, 0.486)
sim <- c(-1.463, 0.027, -2.053, -1.091, -1.602, 2.018, 0.723, -0.776,
         -2.351, 1.054)
wts <- c(-0.05217, 0.08363, 0.07840, -0.00753, -7.35675, -0.00066)
nodes <- c(3, 1, 1)
results <- validann(obs = obs, sim = sim, wts = wts, nodes = nodes)