Going.

fastlib/trunk/contrib/dongryel/thesis_research/mlpack/mixed_logit_dcm/dcm_table.h

@@ -127,6 +127,13 @@ class DCMTable {
 
   public:
 
+    /** @brief Returns the number of attributes for a given discrete
+     *         choice.
+     */
+    int num_attributes() const {
+      return attribute_table_->n_attributes();
+    }
+
     /** @brief Returns the number of discrete choices available for
      *         the given person.
      */

fastlib/trunk/contrib/dongryel/thesis_research/mlpack/mixed_logit_dcm/mixed_logit_dcm_distribution.h

@@ -22,13 +22,53 @@ class MixedLogitDCMDistribution {
     core::table::DensePoint parameters_;
 
   public:
-    virtual double MixedLogitParameterGradient(
+
+    /** @brief Returns the (row, col)-th entry of
+     *         $\frac{\partial}{\partial \theta} \beta^{\nu}(\theta)$
+     */
+    virtual double AttributeGradientWithRespectToParameter(
       int row_index, int col_index) const = 0;
 
     virtual int num_parameters() const = 0;
 
     virtual void Init(const std::string &file_name) const = 0;
 
+    void ChoiceProbabilityWeightedAttributeVector(
+      DCMTableType *dcm_table_in, int person_index,
+      const core::table::DensePoint &choice_probabilities,
+      core::table::DensePoint *choice_prob_weighted_attribute_vector) const {
+
+      // Get the number of discrete choices for the given person.
+      int num_discrete_choices = dcm_table_in->num_discrete_choices(
+                                   person_index);
+      choice_prob_weighted_attribute_vector->Init(
+        dcm_table_in->num_attributes());
+      choice_prob_weighted_attribute_vector->SetZero();
+      for(int i = 0; i < choice_probabilities.length(); i++) {
+        core::table::DensePoint attribute_vector;
+        dcm_table_in->get_attribute_vector(person_index, i);
+        core::math::AddExpert(
+          choice_probabilities[i], attribute_vector,
+          choice_prob_weighted_attribute_vector);
+      }
+    }
+
+    /** @brief Returns the (row, col)-th entry of
+     *         $\frac{\partial}{\partial \beta^2} P_{i j_i^*} (
+     *         \beta^{\nu}(\theta))$ (Equation 8.5)
+     */
+    double HessianChoiceProbability(
+      int row_index, int col_index,
+      const core::table::DensePoint &choice_probabilities,
+      int discrete_choice_index) const {
+
+      double choice_probability = choice_probabilities[discrete_choice_index];
+      double unnormalized_entry = 0;
+
+      // Scale the entry by the choice probability.
+      return choice_probability * unnormalized_entry;
+    }
+
     /** @brief Computes the required quantities in Equation 8.14 (see
      *         dcm_table.h) for a realization of $\beta$ for a given
      *         person.
@@ -41,6 +81,13 @@ class MixedLogitDCMDistribution {
       core::table::DenseMatrix *hessian_first_part,
       core::table::DensePoint *hessian_second_part) const {
 
+      // Intialize the output matrices.
+      hessian_first_part->Init(
+        this->num_parameters(), this->num_parameters());
+      hessian_second_part->Init(this->num_parameters());
+      hessian_first_part->SetZero();
+      hessian_second_part->SetZero();
+
 
     }
 
@@ -49,7 +96,7 @@ class MixedLogitDCMDistribution {
      *         \beta^{\nu}(\theta))$ for a realization of $\beta$ for
      *         a given person.
      */
-    void MixedLogitParameterGradientProducts(
+    void ProductAttributeGradientWithRespectToParameter(
       DCMTableType *dcm_table_in,
       int person_index, int discrete_choice_index,
       const core::table::DensePoint &parameter_vector,
@@ -65,10 +112,17 @@ class MixedLogitDCMDistribution {
       dcm_table_in->get_attribute_vector(
         person_index, discrete_choice_index, &attribute_vector);
 
-      core::table::DensePoint attribute_vector_sub_choice_probability_vector;
+      // Compute the choice probability weighted attribute
+      // vector. This is $\bar{X}_i \bar{P}_i(\beta)$.
+      core::table::DensePoint choice_prob_weighted_attribute_vector;
+      ChoiceProbabilityWeightedAttributeVector(
+        dcm_table_in, person_index, choice_probabilities,
+        &choice_prob_weighted_attribute_vector);
+
+      core::table::DensePoint attribute_vec_sub_choice_prob_weighted_vec;
       core::math::SubInit(
-        choice_probabilities, attribute_vector,
-        &attribute_vector_sub_choice_probability_vector);
+        choice_prob_weighted_attribute_vector, attribute_vector,
+        &attribute_vec_sub_choice_prob_weighted_vec);
 
       // For each row index of the gradient,
       for(int k = 0; k < this->num_parameters(); k++) {
@@ -76,8 +130,8 @@ class MixedLogitDCMDistribution {
         // For each column index of the gradient,
         double dot_product = 0;
         for(int j = 0; j < attribute_vector.length(); j++) {
-          dot_product += attribute_vector_sub_choice_probability_vector[j] *
-                         this->MixedLogitParameterGradient(k, j);
+          dot_product += attribute_vec_sub_choice_prob_weighted_vec[j] *
+                         this->AttributeGradientWithRespectToParameter(k, j);
         }
         (*product_out)[k] = dot_product;
       }