MATH-1622: Simulated annealing variant of "MultiDirectionalTransform".
diff --git a/commons-math-legacy/src/main/java/org/apache/commons/math4/legacy/optim/nonlinear/scalar/noderiv/MultiDirectionalTransform.java b/commons-math-legacy/src/main/java/org/apache/commons/math4/legacy/optim/nonlinear/scalar/noderiv/MultiDirectionalTransform.java
index 28d6c5d..b6f2a73 100644
--- a/commons-math-legacy/src/main/java/org/apache/commons/math4/legacy/optim/nonlinear/scalar/noderiv/MultiDirectionalTransform.java
+++ b/commons-math-legacy/src/main/java/org/apache/commons/math4/legacy/optim/nonlinear/scalar/noderiv/MultiDirectionalTransform.java
@@ -26,7 +26,7 @@
import org.apache.commons.math4.legacy.optim.PointValuePair;
/**
- * Multi-directional search method.
+ * <a href="https://scholarship.rice.edu/handle/1911/16304">Multi-directional</a> search method.
*/
public class MultiDirectionalTransform
implements Simplex.TransformFactory {
@@ -71,7 +71,7 @@
@Override
public UnaryOperator<Simplex> create(final MultivariateFunction evaluationFunction,
final Comparator<PointValuePair> comparator,
- final DoublePredicate unused) {
+ final DoublePredicate sa) {
return original -> {
final PointValuePair best = original.get(0);
@@ -90,9 +90,19 @@
evaluationFunction);
final PointValuePair expandedBest = expandedSimplex.get(0);
- return comparator.compare(reflectedBest, expandedBest) < 0 ?
- reflectedSimplex :
- expandedSimplex;
+ if (comparator.compare(expandedBest, reflectedBest) < 0) {
+ return expandedSimplex;
+ } else if (sa != null &&
+ sa.test(expandedBest.getValue() - reflectedBest.getValue())) {
+ // SA hybridation branch (not part of Torczon's algorithm):
+ // Create a simplex that contains
+ // * the reflected simplex's best point (since it is the
+ // best point overall), and
+ // * the expanded simplex's points (except its best point).
+ return reflectedSimplex.replaceLast(expandedSimplex.asList(1, expandedSimplex.getSize()));
+ } else {
+ return reflectedSimplex;
+ }
} else {
// Compute the contracted simplex.
return original.shrink(sigma, evaluationFunction);
