/**
 * @file nguyen_widrow_init.hpp
 * @author Marcus Edel
 *
 * Definition and implementation of the Nguyen-Widrow method. This
 * initialization rule initialize the weights so that the active regions of the
 * neurons are approximately evenly distributed over the input space.
 *
 * For more information, see the following paper.
 *
 * @code
 * @inproceedings{NguyenIJCNN1990,
 *   title={Improving the learning speed of 2-layer neural networks by choosing
 *   initial values of the adaptive weights},
 *   booktitle={Neural Networks, 1990., 1990 IJCNN International Joint
 *   Conference on},
 *   year={1990}
 * }
 * @endcode
 *
 * mlpack is free software; you may redistribute it and/or modify it under the
 * terms of the 3-clause BSD license.  You should have received a copy of the
 * 3-clause BSD license along with mlpack.  If not, see
 * http://www.opensource.org/licenses/BSD-3-Clause for more information.
 */
#ifndef MLPACK_METHODS_ANN_INIT_RULES_NGUYEN_WIDROW_INIT_HPP
#define MLPACK_METHODS_ANN_INIT_RULES_NGUYEN_WIDROW_INIT_HPP

#include <mlpack/prereqs.hpp>

#include "init_rules_traits.hpp"
#include "random_init.hpp"

namespace mlpack {
namespace ann /** Artificial Neural Network. */ {

/**
 * This class is used to initialize the weight matrix with the Nguyen-Widrow
 * method. The method is defined by
 *
 * @f{eqnarray*}{
 * \gamma &\le& w_i \le \gamma \\
 * \beta &=& 0.7H^{\frac{1}{I}} \\
 * n &=& \sqrt{\sum_{i=0}{I}w_{i}^{2}} \\
 * w_i &=& \frac{\beta w_i}{n}
 * @f}
 *
 * Where H is the number of neurons in the outgoing layer, I represents the
 * number of neurons in the ingoing layer and gamma defines the random interval
 * that is used to initialize the weights with a random value in a specific
 * range.
 */
class NguyenWidrowInitialization
{
 public:
  /**
   * Initialize the random initialization rule with the given lower bound and
   * upper bound.
   *
   * @param lowerBound The number used as lower bound.
   * @param upperBound The number used as upper bound.
   */
  NguyenWidrowInitialization(const double lowerBound = -0.5,
                             const double upperBound = 0.5) :
      lowerBound(lowerBound), upperBound(upperBound) { }

  /**
   * Initialize the elements of the specified weight matrix with the
   * Nguyen-Widrow method.
   *
   * @param W Weight matrix to initialize.
   * @param rows Number of rows.
   * @param cols Number of columns.
   */
  template<typename eT>
  void Initialize(arma::Mat<eT>& W, const size_t rows, const size_t cols)
  {
    RandomInitialization randomInit(lowerBound, upperBound);
    randomInit.Initialize(W, rows, cols);

    double beta = 0.7 * std::pow(cols, 1 / rows);
    W *= (beta / arma::norm(W));
  }

  /**
   * Initialize the elements of the specified weight 3rd order tensor with the
   * Nguyen-Widrow method.
   *
   * @param W Weight matrix to initialize.
   * @param rows Number of rows.
   * @param cols Number of columns.
   * @param slices Number of slices.
   */
  template<typename eT>
  void Initialize(arma::Cube<eT>& W,
                  const size_t rows,
                  const size_t cols,
                  const size_t slices)
  {
    W = arma::Cube<eT>(rows, cols, slices);

    for (size_t i = 0; i < slices; i++)
      Initialize(W.slice(i), rows, cols);
  }

 private:
  //! The number used as lower bound.
  double lowerBound;

  //! The number used as upper bound.
  double upperBound;
}; // class NguyenWidrowInitialization

//! Initialization traits of the Nguyen-Widrow initialization rule.
template<>
class InitTraits<NguyenWidrowInitialization>
{
 public:
  //! The Nguyen-Widrow initialization rule is applied over the entire network.
  static const bool UseLayer = false;
};


} // namespace ann
} // namespace mlpack

#endif