Add test ensuring profiles are initialized respecting their ordering

MICS-Lab · Oct 16, 2023 · fec0731 · fec0731
1 parent 78cfc81
commit fec0731
Showing 1 changed file with 44 additions and 13 deletions.
diff --git a/...es-breed/initialize-profiles/probabilistic-maximal-discrimination-power-per-criterion.cpp b/...es-breed/initialize-profiles/probabilistic-maximal-discrimination-power-per-criterion.cpp
@@ -3,17 +3,22 @@
 #include "probabilistic-maximal-discrimination-power-per-criterion.hpp"
 
 #include "../../../chrones.hpp"
+#include "../../../generation.hpp"  // Only for tests
+
+#include "../../../vendored/doctest.h"  // Keep last because it defines really common names like CHECK that we don't want injected into other headers
 
 
 namespace lincs {
 
 InitializeProfilesForProbabilisticMaximalDiscriminationPowerPerCriterion::InitializeProfilesForProbabilisticMaximalDiscriminationPowerPerCriterion(LearningData& learning_data_) : learning_data(learning_data_) {
   CHRONE();
 
+  generators.reserve(learning_data.criteria_count);
   for (unsigned criterion_index = 0; criterion_index != learning_data.criteria_count; ++criterion_index) {
-    generators.emplace_back();
+    auto& generator = generators.emplace_back();
+    generator.reserve(learning_data.categories_count - 1);
     for (unsigned profile_index = 0; profile_index != learning_data.categories_count - 1; ++profile_index) {
-      generators.back().emplace_back(get_candidate_probabilities(criterion_index, profile_index));
+      generator.emplace_back(get_candidate_probabilities(criterion_index, profile_index));
     }
   }
 }
@@ -88,26 +93,16 @@ void InitializeProfilesForProbabilisticMaximalDiscriminationPowerPerCriterion::i
         const unsigned profile_index = category_index - 1;
         float value = generators[criterion_index][profile_index](learning_data.urbgs[model_index]);
 
+        // Enforce profiles ordering constraint
         if (criterion.category_correlation == Criterion::CategoryCorrelation::growing) {
           if (profile_index != learning_data.categories_count - 2) {
             value = std::min(value, learning_data.profiles[criterion_index][profile_index + 1][model_index]);
           }
-          // @todo(Project management, soon) Add a unit test that triggers the following assertion
-          // (This will require removing the code to enforce the order of profiles above)
-          // Then restore the code to enforce the order of profiles
-          // Note, this assertion does not protect us from initializing a model with two identical profiles.
-          // Is it really that bad?
-          assert(
-            profile_index == learning_data.categories_count - 2
-            || learning_data.profiles[criterion_index][profile_index + 1][model_index] >= value);
         } else {
           assert(criterion.category_correlation == Criterion::CategoryCorrelation::decreasing);
           if (profile_index != learning_data.categories_count - 2) {
             value = std::max(value, learning_data.profiles[criterion_index][profile_index + 1][model_index]);
           }
-          assert(
-            profile_index == learning_data.categories_count - 2
-            || learning_data.profiles[criterion_index][profile_index + 1][model_index] <= value);
         }
 
         learning_data.profiles[criterion_index][profile_index][model_index] = value;
@@ -116,4 +111,40 @@ void InitializeProfilesForProbabilisticMaximalDiscriminationPowerPerCriterion::i
   }
 }
 
+TEST_CASE("Initialize profiles - respect ordering") {
+  Problem problem{
+    {
+      Criterion(
+        "Criterion 1",
+        Criterion::ValueType::real,
+        Criterion::CategoryCorrelation::growing,
+        0.0, 1.0
+      ),
+      Criterion(
+        "Criterion 2",
+        Criterion::ValueType::real,
+        Criterion::CategoryCorrelation::decreasing,
+        0.0, 1.0
+      )
+    },
+    {
+      Category("Category 1"),
+      Category("Category 2"),
+      Category("Category 3"),
+    }
+  };
+  Model model = generate_mrsort_classification_model(problem, 42);
+  auto learning_set = generate_classified_alternatives(problem, model, 1000, 42, 0.1);
+  auto learning_data = LearnMrsortByWeightsProfilesBreed::LearningData::make(problem, learning_set, 1, 42);
+  InitializeProfilesForProbabilisticMaximalDiscriminationPowerPerCriterion initializer(learning_data);
+
+  for (unsigned iteration = 0; iteration != 10; ++iteration) {
+    initializer.initialize_profiles(0, 1);
+    // Both CHECKs fail at least once when the 'Enforce profiles ordering constraint' code is removed
+    CHECK(learning_data.profiles[0][0][0] <= learning_data.profiles[0][1][0]);
+    CHECK(learning_data.profiles[1][0][0] >= learning_data.profiles[1][1][0]);
+  }
+}
+
+
 }  // namespace lincs