Passing Extra Data Into a Function - c++

Question

I am using the dlib optimization library for C++ specifically the following function:
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T
>
double find_max (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
double max_f
);
Functions f and der are designed to take a vector of the data parameters being modified to obtain the maximum value of my function. However my function being maximized has four parameters (one is my dataset and the other is fixed by me). However I can't pass these as inputs to my f and der functions because of the format they are supposed to have. How do I get this data into my functions? I am currently trying the below (I hard set variable c but for vector xgrequ I read data from a file each time I process the function.
//Function to be minimized
double mleGPD(const column_vector& p)
{
std::ifstream infile("Xm-EVT.csv");
long double swapRet;
std::string closeStr;
std::vector<double> histRet;
//Read in historical swap data file
if (infile.is_open())
{
while (!infile.eof())
{
infile >> swapRet;
histRet.push_back(swapRet);
}
}
sort(histRet.begin(), histRet.end());
std::vector<double> negRet;
//separate out losses
for (unsigned c = 0; c < histRet.size(); c++)
{
if (histRet[c] < 0)
{
negRet.push_back(histRet[c]);
}
}
std::vector<double> absValRet;
//make all losses positive to fit with EVT convention
for (unsigned s = 0; s < negRet.size(); s++)
{
absValRet.push_back(abs(negRet[s]));
}
std::vector<double> xminusu, xmu, xgrequ;
int count = absValRet.size();
double uPercent = .9;
int uValIndex = ceil((1 - uPercent)*count);
int countAbove = count - uValIndex;
double c = (double)absValRet[uValIndex - 1];
//looking at returns above u
for (unsigned o = 0; o < uValIndex; ++o)
{
xmu.push_back(absValRet[o] - c);
if (xmu[o] >= 0)
{
xgrequ.push_back(absValRet[o]);
xminusu.push_back(xmu[o]);
}
}
double nu = xgrequ.size();
double sum = 0.0;
double a = p(0);
double b = p(1);
for (unsigned h = 0; h < nu; ++h)
{
sum += log((1 / b)*pow(1 - a*((xgrequ[h] - c) / b), -1 + (1 / a)));
}
return sum;
}
//Derivative of function to be minimized
const column_vector mleGPDDer(const column_vector& p)
{
std::ifstream infile("Xm-EVT.csv");
long double swapRet;
std::string closeStr;
std::vector<double> histRet;
//Read in historical swap data file
if (infile.is_open())
{
while (!infile.eof())
{
infile >> swapRet;
histRet.push_back(swapRet);
}
}
sort(histRet.begin(), histRet.end());
std::vector<double> negRet;
//separate out losses
for (unsigned c = 0; c < histRet.size(); c++)
{
if (histRet[c] < 0)
{
negRet.push_back(histRet[c]);
}
}
std::vector<double> absValRet;
//make all losses positive to fit with EVT convention
for (unsigned s = 0; s < negRet.size(); s++)
{
absValRet.push_back(abs(negRet[s]));
}
std::vector<double> xminusu, xmu, xgrequ;
int count = absValRet.size();
double uPercent = .9;
int uValIndex = ceil((1 - uPercent)*count);
int countAbove = count - uValIndex;
double c = (double)absValRet[uValIndex - 1];
//looking at returns above u
for (unsigned o = 0; o < uValIndex; ++o)
{
xmu.push_back(absValRet[o] - c);
if (xmu[o] >= 0)
{
xgrequ.push_back(absValRet[o]);
xminusu.push_back(xmu[o]);
}
}
column_vector res(2);
const double a = p(0);
const double b = p(1);
double nu = xgrequ.size();
double sum1 = 0.0;
double sum2 = 0.0;
for (unsigned h = 0; h < nu; ++h)
{
sum1 += ((xgrequ[h]-c)/b)/(1-a*((xgrequ[h]-c)/b));
sum2 += log(1 - a*((xgrequ[h] - c) / b));
}
res(0) = sum1;//df/da
res(1) = sum2;//df/db
return res;
}
Here is what my actual function call looks like:
//Dlib max finding
column_vector start(2);
start = .1, .1; //starting point for a and b
find_max(bfgs_search_strategy(), objective_delta_stop_strategy(1e-6), mleGPD, mleGPDDer, start,100);
std::cout << "solution" << start << std::endl;

Answer 1

This kind of API is very common. It's almost always possible to for f and der to any callable, not just static functions. that is, you can pass a custom class object with operator () to it.
For example
struct MyF {
//int m_state;
// or other state variables, such as
std::vector<double> m_histRet;
// (default constructors will do)
double operator()(const column_vector& p) const {
return some_function_of(p, m_state);
}
};
int main(){
. . .
MyF myf{42};
// or
MyF myf{someVectorContainingHistRet};
// then use myf as you would have used mleGPD
}
You need to initiate MyF and MyDer with the same state (std::vector<double> histRet I presume.) Either as copies or (const) references to the same state.
Edit: More full example:
struct MLGDPG_State {
std::vector<double> xgrequ;
// . . . and more you need in f or fder
}
MLGDPG_State makeMLGDPG_State(const std::string& filename){
std::ifstream infile(filename);
std::ifstream infile("Xm-EVT.csv");
long double swapRet;
std::string closeStr;
std::vector<double> histRet;
//Read in historical swap data file
if (infile.is_open())
{
while (!infile.eof())
{
infile >> swapRet;
histRet.push_back(swapRet);
}
}
sort(histRet.begin(), histRet.end());
std::vector<double> negRet;
//separate out losses
for (unsigned c = 0; c < histRet.size(); c++)
{
if (histRet[c] < 0)
{
negRet.push_back(histRet[c]);
}
}
std::vector<double> absValRet;
//make all losses positive to fit with EVT convention
for (unsigned s = 0; s < negRet.size(); s++)
{
absValRet.push_back(abs(negRet[s]));
}
std::vector<double> xminusu, xmu, xgrequ;
int count = absValRet.size();
double uPercent = .9;
int uValIndex = ceil((1 - uPercent)*count);
int countAbove = count - uValIndex;
double c = (double)absValRet[uValIndex - 1];
//looking at returns above u
for (unsigned o = 0; o < uValIndex; ++o)
{
xmu.push_back(absValRet[o] - c);
if (xmu[o] >= 0)
{
xgrequ.push_back(absValRet[o]);
xminusu.push_back(xmu[o]);
}
}
return {std::move(xgrequ)};
// Or just 'return MleGPD(xgrequ)' if you are scared of {} and move
}
//Functor Class, for ion to be minimized
struct MleGPD{
MLGDPG_State state;
double operator()(const column_vector& p) const {
auto mu = state.xgrequ.size();
double sum = 0.0;
double a = p(0);
double b = p(1);
for (unsigned h = 0; h < nu; ++h)
{
sum += log((1 / b)*pow(1 - a*((xgrequ[h] - c) / b), -1 + (1 / a)));
}
return sum;
};
Use the same pattern for a struct MleGPD_Derivative.
Usage:
const auto state = makeMLGDPG_State("Xm-EVT.csv");
const auto f = MleGPD{state};
const auto der = MleGPD_Derivative{state};
start = .1, .1; //starting point for a and b
find_max(bfgs_search_strategy(), objective_delta_stop_strategy(1e-6), f, der, start,100);
std::cout << "solution" << start << std::endl;
Note, that for simple structs like these, it's often fine to use the default constructors, copy constructor etc. Also note http://en.cppreference.com/w/cpp/language/aggregate_initialization