More brainstorming, back to distributed kde.

2010-12-14 21:01:31 +00:00
parent bc8bcc3f44
commit ecfbfacf97
4 changed files with 52 additions and 20 deletions
@@ -12,6 +12,7 @@
 #include "core/math/linear_algebra.h"
 #include "core/monte_carlo/mean_variance_pair.h"
 #include "core/monte_carlo/mean_variance_pair_matrix.h"
+#include "mlpack/mixed_logit_dcm/mixed_logit_dcm_distribution.h"

 namespace mlpack {
 namespace mixed_logit_dcm {
@@ -19,6 +20,11 @@ template<typename TableType>
 class DCMTable {
  private:

+    /** @brief The distribution from which each $\beta$ is sampled
+     *         from.
+     */
+    mlpack::mixed_logit_dcm::MixedLogitDCMDistribution *distribution_;
+
    /** @brief The pointer to the attribute vector for each person per
     *         his/her discrete choice.
     */
@@ -216,19 +222,17 @@ class DCMTable {
      return static_cast<int>(cumulative_num_discrete_choices_.size());
    }

-    void Init(
-      TableType *attribute_table_in,
-      TableType *num_discrete_choices_per_person_in,
-      int num_parameters_in) {
+    template<typename ArgumentType>
+    void Init(ArgumentType &argument_in) {

      // Set the number of parameters.
-      num_parameters_ = num_parameters_in;
+      num_parameters_ = argument_in.distribution_->num_parameters();

      // Set the incoming attributes table and the number of choices
      // per person in the list.
-      attribute_table_ = attribute_table_in;
+      attribute_table_ = argument_in.attribute_table_;
      num_discrete_choices_per_person_.resize(
-        num_discrete_choices_per_person_in->n_entries());
+        argument_in.num_discrete_choices_per_person_->n_entries());

      // This vector maintains the running simulated choice
      // probabilities per person per discrete choice. It is indexed
@@ -250,7 +254,7 @@ class DCMTable {
      // Initialize a randomly shuffled vector of indices for sampling
      // the outer term in the simulated log-likelihood.
      shuffled_indices_for_person_.resize(
-        num_discrete_choices_per_person_in->n_entries());
+        argument_in.num_discrete_choices_per_person_->n_entries());
      for(unsigned int i = 0; i < shuffled_indices_for_person_.size(); i++) {
        shuffled_indices_for_person_[i] = i;
      }
@@ -264,12 +268,12 @@ class DCMTable {
      // index in the attribute table for given (person, discrete
      // choice) pair.
      cumulative_num_discrete_choices_.resize(
-        num_discrete_choices_per_person_in->n_entries());
+        argument_in.num_discrete_choices_per_person_->n_entries());
      cumulative_num_discrete_choices_[0] = 0;
      for(unsigned int i = 1; i < cumulative_num_discrete_choices_.size();
          i++) {
        core::table::DensePoint point;
-        num_discrete_choices_per_person_in->get(i - 1, &point);
+        argument_in.num_discrete_choices_per_person_->get(i - 1, &point);
        int num_choices_for_current_person =
          static_cast<int>(point[0]);
        cumulative_num_discrete_choices_[i] =
@@ -283,14 +287,14 @@ class DCMTable {
      // distribution on the number of choices match up the total
      // number of attribute vectors. Otherwise, quit.
      core::table::DensePoint last_count_vector;
-      num_discrete_choices_per_person_in->get(
+      argument_in.num_discrete_choices_per_person_->get(
        cumulative_num_discrete_choices_.size() - 1, &last_count_vector);
      num_discrete_choices_per_person_[
        cumulative_num_discrete_choices_.size() - 1 ] = last_count_vector[0];
      int last_count = static_cast<int>(last_count_vector[0]);
      if(cumulative_num_discrete_choices_[
            cumulative_num_discrete_choices_.size() - 1] +
-          last_count != attribute_table_in->n_entries()) {
+          last_count != argument_in.attribute_table_->n_entries()) {
        std::cerr << "The total number of discrete choices do not equal "
                  "the number of total number of attribute vectors.\n";
        exit(0);
@@ -316,13 +320,19 @@ class DCMTable {
      ComputeChoiceProbabilities_(
        person_index, parameter_vector, &choice_probabilities);

+      // Loop through and update the simulated choice probabilities
+      // and the simulated log-likelihood gradients.
      int index = cumulative_num_discrete_choices_[person_index];
      for(int num_discrete_choices = 0; num_discrete_choices <
          num_discrete_choices_per_person_[person_index];
          num_discrete_choices++, index++) {

+        // Simulated choice probability update.
        simulated_choice_probabilities_[index].push_back(
          choice_probabilities[num_discrete_choices]);
+
+        // Simulated log-likelihood gradient update.
+        // simulated_loglikelihood_gradients_[index].push_back();
      }
    }

@@ -7,6 +7,7 @@
 #define MLPACK_MIXED_LOGIT_DCM_MIXED_LOGIT_DCM_ARGUMENTS_H

 #include "core/table/table.h"
+#include "mlpack/mixed_logit_dcm/mixed_logit_dcm_distribution.h"

 namespace mlpack {
 namespace mixed_logit_dcm {
@@ -17,7 +18,7 @@ class MixedLogitDCMArguments {

    TableType *num_discrete_choices_per_person_;

-    int num_parameters_;
+    mlpack::mixed_logit_dcm::MixedLogitDCMDistribution *distribution_;

    double initial_dataset_sample_rate_;

@@ -32,7 +33,7 @@ class MixedLogitDCMArguments {
    MixedLogitDCMArguments() {
      attribute_table_ = NULL;
      num_discrete_choices_per_person_ = NULL;
-      num_parameters_ = 0;
+      distribution_ = NULL;
      initial_dataset_sample_rate_ = 0;
      initial_integration_sample_rate_ = 0;
    }
@@ -42,6 +43,8 @@ class MixedLogitDCMArguments {
      attribute_table_ = NULL;
      delete num_discrete_choices_per_person_;
      num_discrete_choices_per_person_ = NULL;
+      delete distribution_;
+      distribution_ = NULL;
    }
 };
 };
@@ -18,10 +18,7 @@ void MixedLogitDCM<TableType>::Init(

  // Initialize the table for storing/accessing the attribute vector
  // for each person.
-  table_.Init(
-    arguments_in.attribute_table_,
-    arguments_in.num_discrete_choices_per_person_,
-    arguments_in.num_parameters_);
+  table_.Init(arguments_in);
 }

 template<typename TableType>
@@ -173,8 +170,7 @@ void MixedLogitDCM<TableType>::ParseArguments(

  // The number of parameters that generate each $\beta$ is fixed now
  // as the Gaussian example in Appendix.
-  arguments_out->num_parameters_ = 2 *
-                                   arguments_out->attribute_table_->n_attributes();
+  // arguments_out->distribution_ = ;
 }

 template<typename TableType>
@@ -0,0 +1,23 @@
+/** @file mixed_logit_dcm_distribution.h
+ *
+ *  @author Dongryeol Lee (dongryel@cc.gatech.edu)
+ */
+
+#ifndef MLPACK_MIXED_LOGIT_DCM_DCM_DISTRIBUTION_H
+#define MLPACK_MIXED_LOGIT_DCM_DCM_DISTRIBUTION_H
+
+namespace mlpack {
+namespace mixed_logit_dcm {
+
+/** @brief The base abstract class for the distribution that
+ *         generates each $\beta$ parameter in mixed logit
+ *         models. This distribution is parametrized by $\theta$.
+ */
+class MixedLogitDCMDistribution {
+  public:
+    virtual int num_parameters() const = 0;
+};
+};
+};
+
+#endif