Prev Next optimize_forward_active.cpp Headings

@(@\newcommand{\W}[1]{ \; #1 \; } \newcommand{\R}[1]{ {\rm #1} } \newcommand{\B}[1]{ {\bf #1} } \newcommand{\D}[2]{ \frac{\partial #1}{\partial #2} } \newcommand{\DD}[3]{ \frac{\partial^2 #1}{\partial #2 \partial #3} } \newcommand{\Dpow}[2]{ \frac{\partial^{#1}}{\partial {#2}^{#1}} } \newcommand{\dpow}[2]{ \frac{ {\rm d}^{#1}}{{\rm d}\, {#2}^{#1}} }@)@
Example Optimization and Forward Activity Analysis
# include <cppad/cppad.hpp>
namespace {
     struct tape_size { size_t n_var; size_t n_op; };

     template <class Vector> void fun(
          const Vector& x, Vector& y, tape_size& before, tape_size& after
     )
     {     typedef typename Vector::value_type scalar;

          // phantom variable with index 0 and independent variables
          // begin operator, independent variable operators and end operator
          before.n_var = 1 + x.size(); before.n_op  = 2 + x.size();
          after.n_var  = 1 + x.size(); after.n_op   = 2 + x.size();

          // adding the constant zero does not take any operations
          scalar zero   = 0.0 + x[0];
          before.n_var += 0; before.n_op  += 0;
          after.n_var  += 0; after.n_op   += 0;

          // multiplication by the constant one does not take any operations
          scalar one    = 1.0 * x[1];
          before.n_var += 0; before.n_op  += 0;
          after.n_var  += 0; after.n_op   += 0;

          // multiplication by the constant zero does not take any operations
          // and results in the constant zero.
          scalar two    = 0.0 * x[0];

          // operations that only involve constants do not take any operations
          scalar three  = (1.0 + two) * 3.0;
          before.n_var += 0; before.n_op  += 0;
          after.n_var  += 0; after.n_op   += 0;

          // The optimizer will reconize that zero + one = one + zero
          // for all values of x.
          scalar four   = zero + one;
          scalar five   = one  + zero;
          before.n_var += 2; before.n_op  += 2;
          after.n_var  += 1; after.n_op   += 1;

          // The optimizer will reconize that sin(x[3]) = sin(x[3])
          // for all values of x. Note that, for computation of derivatives,
          // sin(x[3]) and cos(x[3]) are stored on the tape as a pair.
          scalar six    = sin(x[2]);
          scalar seven  = sin(x[2]);
          before.n_var += 4; before.n_op  += 2;
          after.n_var  += 2; after.n_op   += 1;

          // If we used addition here, five + seven = zero + one + seven
          // which would get converted to a cumulative summation operator.
          scalar eight = five * seven;
          before.n_var += 1; before.n_op  += 1;
          after.n_var  += 1; after.n_op   += 1;

          // Use two, three, four and six in order to avoid a compiler warning
          // Note that addition of two and three does not take any operations.
          // Also note that optimizer reconizes four * six == five * seven.
          scalar nine  = eight + four * six * (two + three);
          before.n_var += 3; before.n_op  += 3;
          after.n_var  += 2; after.n_op   += 2;

          // results for this operation sequence
          y[0] = nine;
          before.n_var += 0; before.n_op  += 0;
          after.n_var  += 0; after.n_op   += 0;
     }
}

bool forward_active(void)
{     bool ok = true;
     using CppAD::AD;
     using CppAD::NearEqual;
     double eps10 = 10.0 * std::numeric_limits<double>::epsilon();

     // domain space vector
     size_t n  = 3;
     CPPAD_TESTVECTOR(AD<double>) ax(n);
     ax[0] = 0.5;
     ax[1] = 1.5;
     ax[2] = 2.0;

     // declare independent variables and start tape recording
     CppAD::Independent(ax);

     // range space vector
     size_t m = 1;
     CPPAD_TESTVECTOR(AD<double>) ay(m);
     tape_size before, after;
     fun(ax, ay, before, after);

     // create f: x -> y and stop tape recording
     CppAD::ADFun<double> f(ax, ay);
     ok &= f.size_var() == before.n_var;
     ok &= f.size_op()  == before.n_op;

     // Optimize the operation sequence
     // Note that, for this case, all the optimization was done during
     // the recording and there is no benifit to the optimization.
     f.optimize();
     ok &= f.size_var() == after.n_var;
     ok &= f.size_op()  == after.n_op;

     // check zero order forward with different argument value
     CPPAD_TESTVECTOR(double) x(n), y(m), check(m);
     for(size_t i = 0; i < n; i++)
          x[i] = double(i + 2);
     y    = f.Forward(0, x);
     fun(x, check, before, after);
     ok &= NearEqual(y[0], check[0], eps10, eps10);

     return ok;
}

Input File: example/optimize/forward_active.cpp