[ML] Fix potential use of unintialised value when minimising lowess r…

…egression function (#2268)

If for any reason the extrapolation range for lowess regression returns zero then globalMinimize, to find the function
minimum, would be supplied an empty collection of points and fail. This would then lead to minimum returning a
uninitialised value for the minimum. This shouldn't normally happen in our usage of it, but we have seen the error
message associated with supplying an empty collection to globalMinimize in our QA suite.

docs/CHANGELOG.asciidoc

@@ -46,6 +46,8 @@
   classification and regression models. (See {ml-pull}2259[#2259].)
 * Fix edge case which could cause the model bounds to blow up after detecting seasonality.
   (See {ml-pull}2261[#2261].)
+* Fix cause of "Must provide points at which to evaluate function" log error training
+  classification and regression models. (See {ml-pull}2268[#2268].)
 
 == {es} version 8.2.0
 

include/maths/common/CLowessDetail.h

@@ -152,13 +152,17 @@ typename CLowess<N>::TDoubleDoublePr CLowess<N>::minimum() const {
     double range{(xb - xa) / static_cast<double>(X.size())};
     xa = std::max(xa, xmin - 0.5 * range);
     xb = std::min(xb, xmin + 0.5 * range);
+    if (xa == xb) {
+        return {xmin, fmin};
+    }
+
     double dx{2.0 * (xb - xa) / static_cast<double>(X.size())};
     X.clear();
     for (double x = xa; x < xb; x += dx) {
         X.push_back(x);
     }
-    double xcand;
-    double fcand;
+    double xcand{xmin};
+    double fcand{fmin};
     CSolvers::globalMinimize(
         X, [this](double x) -> double { return this->predict(x); }, xcand, fcand, fsd);