train.ann: Training an Artificial Neural Network for Interaction...

In quarrint: Interaction Prediction Between Groundwater and Quarry Extension Using Discrete Choice Models and Artificial Neural Networks

Description

The function trains a neural network to be used with the functions compute.interaction and compute.ann. The neural network can then be used to predict whether the level of interaction between a quarry and the groundwater is low, medium, high or very high.

The user can specify:

• the explanatory variables to be used;

• the data frame used to train and validate the network;

• the structure of the hidden layers;

• the number of repetitions for the neural network training.

Usage

train.ann(var = c("H", "Z", "G", "C", "T", "L"), data = quarrint::quarries, 
          hidden = 7, rep = 1, ...)

Arguments

var	The explanatory variable to be used. By default, all the variables in the default data frame are used. Note that the variables must be categorical (coded with integers) and will be transformed in dummy variables. For instance if X has 5 possible values (1,2,3,4,5), then it will be replaced by the binary variables X1, X2, X3, X4, X5.
data	The training and validation dataframe. It must contain the variables listed in "var" and the dependent binary variables "low", "medium", "high" and "very.high" representing the interaction level. By default, it uses the dataframe "quarries" from the package.
hidden	A vector of integer detailing the structure of the hidden layers. For instance if we want 2 hidden layers with 4 and 2 nodes respectively, then it must be it to (2, 4). The default is 7, i.e. 1 hidden layer of 7 nodes.
rep	The number of repetition of the neural network to be computed.
...	Further arguments passed to or from other methods. See the documentation of "neuralnet" from the package "neuralnet".

Value

A list whose elements are:

ann	A nn object containing the trained neural network.
prop.correct.prediction	A list detailing for each repetition of the neural network the proportion of correct predictions.

Author(s)

Johan Barthelemy.

Maintainer: Johan Barthelemy [e-mail].

References

Barthelemy, J., Carletti, T., Collier L., Hallet, V., Moriame, M., Sartenaer, A. (2016) Interaction prediction between groundwater and quarry extension using discrete choice models and artificial neural networks Environmental Earth Sciences (in press)

Collier, L., Barthelemy, J., Carletti, T., Moriame, M., Sartenaer, A., Hallet, V. (2015) Calculation of an Interaction Index between the Extractive Activity and Groundwater Resources Energy Procedia 76, 412-420

Krieselm, D. (2007) A Brief Introduction to Neural Networks. On-line available at http://www.dkriesel.com

Ripley, B. (1996) Pattern recognition and neural networks Cambridge university press

See Also

The function relies on the function neuralnet of the neuralnet package to generate an object of type nn containing the trained neural network.

compute.interaction and compute.ann to use the trained neural network.

The data frame quarries.

Examples

## Not run: 
# training a neural network using the attribues H and T as predictors,
# with 2 hidden layers of 2 nodes each and computing 2 replications
r.ann <- train.ann(var = c("H", "T", "L", "Z"), hidden = c(2, 2), rep = 2)

# using the ann to compute the interaction
q <- quarry(geological.context = 2, hydrogeological.context = 4,
            piezometric.context = 1, quarry.position = 4,
            production.catchment = 4, quality.catchment = 3)
compute.interaction(q, method="ann", ann = r.ann$ann)

## End(Not run)

quarrint documentation built on May 1, 2019, 10:10 p.m.