[ML] Apply tree depth constraint to hyperparameter search bounding box setup (#1870)

Following on from #1867, we can and should be imposing the minimum depth constraint to the hyperparameter search
bounding box. This was incorrectly applied before and also fixes the issue with reproducibility based on user overrides.
This is a bit cleaner than applying the constraint magically in the code to adjust hyperparameters.

Author: tveasey

lib/maths/CBoostedTreeFactory.cc

@@ -570,8 +570,8 @@ void CBoostedTreeFactory::initializeUnsetRegularizationHyperparameters(core::CDa
                                                  -mainLoopSearchInterval / 2.0,
                                                  mainLoopSearchInterval / 2.0)
                             .value_or(fallback);
-                    m_SoftDepthLimitSearchInterval =
-                        max(m_SoftDepthLimitSearchInterval, TVector{1.0});
+                    m_SoftDepthLimitSearchInterval = max(
+                        m_SoftDepthLimitSearchInterval, TVector{MIN_SOFT_DEPTH_LIMIT});
                     LOG_TRACE(<< "soft depth limit search interval = ["
                               << m_SoftDepthLimitSearchInterval.toDelimited() << "]");
                     m_TreeImpl->m_Regularization.softTreeDepthLimit(

lib/maths/CBoostedTreeImpl.cc

@@ -1417,7 +1417,7 @@ bool CBoostedTreeImpl::selectNextHyperparameters(const TMeanVarAccumulator& loss
                 scale * CTools::stableExp(parameters(i)));
             break;
         case E_SoftTreeDepthLimit:
-            m_Regularization.softTreeDepthLimit(std::max(parameters(i), 2.0));
+            m_Regularization.softTreeDepthLimit(parameters(i));
             break;
         case E_SoftTreeDepthTolerance:
             m_Regularization.softTreeDepthTolerance(parameters(i));

lib/maths/unittest/CBoostedTreeTest.cc

@@ -935,7 +935,7 @@ BOOST_AUTO_TEST_CASE(testCategoricalRegressors) {
     LOG_DEBUG(<< "bias = " << modelBias);
     LOG_DEBUG(<< " R^2 = " << modelRSquared);
     BOOST_REQUIRE_CLOSE_ABSOLUTE(0.0, modelBias, 0.16);
-    BOOST_TEST_REQUIRE(modelRSquared > 0.95);
+    BOOST_TEST_REQUIRE(modelRSquared > 0.93);
 }
 
 BOOST_AUTO_TEST_CASE(testFeatureBags) {
@@ -1354,7 +1354,7 @@ BOOST_AUTO_TEST_CASE(testBinomialLogisticRegression) {
         LOG_DEBUG(<< "log relative error = "
                   << maths::CBasicStatistics::mean(logRelativeError));
 
-        BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 0.69);
+        BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 0.70);
         meanLogRelativeError.add(maths::CBasicStatistics::mean(logRelativeError));
     }
 
@@ -1587,7 +1587,7 @@ BOOST_AUTO_TEST_CASE(testMultinomialLogisticRegression) {
         LOG_DEBUG(<< "log relative error = "
                   << maths::CBasicStatistics::mean(logRelativeError));
 
-        BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 2.1);
+        BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 2.2);
         meanLogRelativeError.add(maths::CBasicStatistics::mean(logRelativeError));
     }