Tweaked find_min_box_constrained() so that the user can easily reuse

computations done in f() when computing der().

dlib/optimization/optimization_abstract.h

@@ -330,6 +330,10 @@ namespace dlib
             - returns f(#x). 
             - When calling f() and der(), the input passed to them will always be inside
               the box constraints defined by x_lower and x_upper.
+            - When calling der(x), it will always be the case that the last call to f() was
+              made with the same x value.  This means that you can reuse any intermediate
+              results from the previous call to f(x) inside der(x) rather than recomputing
+              them inside der(x).
     !*/
 
 // ----------------------------------------------------------------------------------------
@@ -419,6 +423,10 @@ namespace dlib
             - returns f(#x). 
             - When calling f() and der(), the input passed to them will always be inside
               the box constraints defined by x_lower and x_upper.
+            - When calling der(x), it will always be the case that the last call to f() was
+              made with the same x value.  This means that you can reuse any intermediate
+              results from the previous call to f(x) inside der(x) rather than recomputing
+              them inside der(x).
             - Note that this function solves the maximization problem by converting it 
               into a minimization problem.  Therefore, the values of f and its derivative
               reported to the stopping strategy will be negated.  That is, stop_strategy

dlib/optimization/optimization_line_search.h

@@ -452,10 +452,12 @@ namespace dlib
         if (d0 > 0 && alpha > 0)
             alpha *= -1;
 
-        for (unsigned long iter = 0; iter < max_iter; ++iter)
+        unsigned long iter = 0;
+        while (true)
         {
+            ++max_iter;
             const double val = f(alpha);
-            if (val <= f0 + alpha*rho*d0)
+            if (val <= f0 + alpha*rho*d0 || iter >= max_iter)
             {
                 return alpha;
             }
@@ -472,7 +474,6 @@ namespace dlib
                 alpha *= step;
             }
         }
-        return alpha;
     }
 
 // ----------------------------------------------------------------------------------------

dlib/optimization/optimization_line_search_abstract.h

@@ -217,10 +217,14 @@ namespace dlib
               rule to decide when the search can stop.
                 - rho == the parameter of the sufficient decrease condition. 
                 - max_iter == the maximum number of iterations allowable.  After this many
-                  evaluations of f() line_search() is guaranteed to terminate.
+                  evaluations of f() backtracking_line_search() is guaranteed to terminate.
             - The line search starts with the input alpha value and then backtracks until
               it finds a good enough alpha value.  Once found, it returns the alpha value
               such that f(alpha) is significantly closer to the minimum of f than f(0).
+            - The returned value of alpha will always be the last value of alpha which was
+              passed to f().  That is, it will always be the case that the last call to f()
+              made by backtracking_line_search() was f(alpha) where alpha is the return
+              value from this function.
     !*/
 
 // ----------------------------------------------------------------------------------------