Utilise Boost to provide some sort of clue for unknown errors

Hi there,

a) Thanks for providing bayesopt -- I am having fun with it and my multidimensional optimisation problems!

b) Please forgive this somewhat specific / longwinded way of asking for a diff to be merged, provided below, for what is undoubtedly technically a Matlab bug. I am providing a vast amount of information in case it helps someone else. On the Matlab (R2023a) interface, I was trying to optimise a 37-dimensional function that returns essentially the least squares difference of a simulation and experimental data. I had defined two functions: (i) optimFun(x) as an anonymous function, and AFullSimulation(x, parameters_1, parameters2, options,experimental_data) differently; that is AFullSimulation(...) was defined separately and optimFun = @(x) (AFullSimulation(x, Sys1, Exp, Options,...) was defined inline in the matlab file I was trying to run. Trying to call bayesoptcont(@optimFun,...) directly failed -- with "Error: -1" returned and ERROR: Unknown error in the log. I tried to understand this and tried iterations such as bayesoptcont('optimFun') etc but to no avail.

I therefore edited the wrapper function's generic, all-purpose catch(...) statement to provide a specific error class if one was available, utilising a feature in boostlib, which you are already linking against:

diff --git a/src/wrappers/bayesoptwpr.cpp b/src/wrappers/bayesoptwpr.cpp
index 2d38591..eeb13d3 100644
--- a/src/wrappers/bayesoptwpr.cpp
+++ b/src/wrappers/bayesoptwpr.cpp
@@ -26,6 +26,7 @@
 #include "log.hpp"
 #include "ublas_extra.hpp"
 #include "specialtypes.hpp"
+#include <boost/exception/diagnostic_information.hpp>

 static const int BAYESOPT_FAILURE = -1; /* generic failure code */
 static const int BAYESOPT_INVALID_ARGS = -2;
@@ -134,7 +135,8 @@ int bayes_optimization(int nDim, eval_func f, void* f_data,
     }
   catch (...)
     {
-      FILE_LOG(logERROR) << "Unknown error";
+      FILE_LOG(logERROR) << "Unknown error building continuous optimizer!" << std::endl;
+      FILE_LOG(logERROR) << boost::current_exception_diagnostic_information() << std::endl; /* Provide debug information if available */
       return BAYESOPT_FAILURE;
     }
   return 0; /* everything ok*/
@@ -192,7 +194,8 @@ int bayes_optimization_disc(int nDim, eval_func f, void* f_data,
     }
   catch (...)
     {
-      FILE_LOG(logERROR) << "Unknown error";
+      FILE_LOG(logERROR) << "Unknown error building discrete optimizer!" << std::endl;
+      FILE_LOG(logERROR) << boost::current_exception_diagnostic_information() << std::endl; /* Provide debugging info if it is there */
       return BAYESOPT_FAILURE;
     }

For my specific Matlab woe, this helpfully prints out the following:

[...] ERROR: Dynamic exception type: foundation::core::except::Exception<MATLAB::legacy_two_part::undefinedFunctionTextInputArgumentsType, std::exception, void> std::exception::what: Undefined function 'optimFun' for input arguments of type 'double'

Replacing @optimFun or 'optimFun' with @(x) AFullSimulation(x,...) (i.e. its full definition writ-large) appears to make Matlab JIT it into something that the mex interface is happy with.

I'm happy to submit this as a pull request if you like but thought I'd reach out first (!).

rmcantin / bayesopt

Utilise Boost to provide some sort of clue for unknown errors #31