Tmp

Putting it all together -- Optimization and EPAnet

In this lesson, we push the abilities we have gained to the limit! We link two code libraries, namely, Evolving objects -- a versatile genetic algorithm code and EPAnet (toolkit) a pipe network calculation model by writing some relatively simple code and create a running program. The following section explains the problem we have selected to solve (please note that the problem itself is of secondary importance here, what we are trying to do is to hone the skills we have gained and build confidence to attack bigger coding projects).

Problem

The problem network.

The figure on the left shows a water supply network with a reservoir, seven pipe segments conveying water and five junctions that have different demands. We want to compute the most economical pipe diameters for each segment while maintaining a minimum pressure head of 10 m at all junctions. You can open the File:Cbcpp pipe network1.inp in Epanet 2 software to view the network.

Our plan is to use a Genetic Algorithm to optimize the pipe diameters. For this we need to connect Epanet Software to a genetic algorithm code.

Plan

We will use the Epanet Toolkit, a programming interface designed to run Epanet 2.0 programmatically without the standard graphical interface. For the Genetic Algorithm part, we'll use Evolving objects a free and open source package for evolutionary computations.

Evolving objects can be downloaded from eodev] website. Epanet toolkit can be downloaded from [US-EPA]. However, for this exercise you may use the already downloaded versions of these tools. Download the file cbcpp_EO.zip into your computer and extract it to a new folder (say E:\GA\EO). This will create four subfolders,

code
data
eo-1.0
epanet_tools

We use code folder to keep all the code we write (not the code libraries we use!). It already have the files:

RealEA.cpp
real_value.h

RealEA.cpp is a c++ program that can be used to access the EO Genetic algorithm functions. real_value.h has the cost function for the GA. We'll learn more about this later.

data folder where we shall keep all our data, has the water distribution network file.

eo-1.0 and epanet_tools have the evolving objects and epanet toolkit code libraries.

Solution

We shall attack our problem in a piecemeal fashion, in the steps given below:

1. Get EO to solve a small GA problem.
2. Replace the cost function with our own.
3. Use epanet_toolkit to run Epanet on our water supply network and use the results to evaluate a cost function.

Usually this step-by-step approach is less error-prone than attempting the whole task once.

Running a GA

Create a new folder called projects under the same folder that has sub-folders of code, data, etc. This is where we shall keep the visual studio express edition related (platform specific) stuff. Open visual C++ and reate a new empty project EPGA in that folder. Add the following files in the code folder to the project.

 
RealEA.cpp
real_value.h

When you try to compile the project, you will get the error message on the line:

#include <eo>

Indicating that the compiler can not include 'eo'. To remedy this, we should include the path to the include file eo ("..\..\eo-1.0.1\src"). This can be done by additing it to: Edit-><project>Properties->C/C++->General->Additional include directories.

At this stage RealEA.cpp should comile successfully, but would cause errors at linking stage. The error messages would look like

unresolved external symbol "public: virtual __thiscall eoFunctorStore::~eoFunctorStore(void)" 
(??1eoFunctorStore@@UAE@XZ) referenced in function 
"public: virtual void * __thiscall eoFunctorStore::`scalar deleting destructor'(unsigned int)" 
(??_GeoFunctorStore@@UAEPAXI@Z)

The reason for this is

1. The compiler knows about eo library (by #include <eo>), but
2. the real library objects of eo needed for linking are missing

To rectify this, we should let the linker access to the necessary libraries. Before doing this we have to make a detour. First save your project/solution and close it.

Building EO libraries

The eo code available for download does not have the libraries for windows built in, but they have the complete source and tools needed to build them. There is a visual C++ solution called eo.sln in the sub-folder win. Just build this solution (Project->Build solution). This will create the necessary libraries in the win.

One more point: It is beneficial to build the release version (not the debug version) of the libraries for performance reasons.

After this you will have the following libraries in the folder win\lib\release.

eo.lib  eoes.lib  eoga.lib  eoutils.lib

Adding linker dependancies.

Now open your project again. In add the above four files as dependancies for the linker. (Edit-><project>Properties->Linker->General->Additional Dependancies).

Then let the linker know where these are: (Edit-><project>Properties->Linker->General->Additional Library Directories). Add something like ..\..\eo-1.0.1\win\lib\release.

At this stage, you should be able to compile the project successfully. Debug->Start without Debugging should run the program, albeit with not-so-meaningful-at-the-moment results.

Enable debugging

If you try to debug your code at this stage, visual c++ 2005 will complain (as of 01:07, 9 April 2007 (JST)) that the binaries were not built with debug information. There is a separate article on how to enable debugging in visual studio 2005.

EO In-Action

Now is a good time to have an idea about how our GA code works. Don't worry if you can not understand everything -- what is important is to have a general idea of how things work!

The actual code-in-action is very short, indeed. I have changed the comments a little bit.

  /* Many instructions to this program can be given on the command line. 
     The following code understands what you have specified as command line arguments. 
  */
  eoParser parser(argc, argv);  // for user-parameter reading
  eoState state;    // keeps all things allocated
  
  typedef eoReal<eoMinimizingFitness> EOT;

  // The evaluation fn - encapsulated into an eval counter for output 
  eoEvalFuncPtr<EOT, double, const std::vector<double>&> 
               mainEval( real_value );
  eoEvalFuncCounter<EOT> eval(mainEval);
  // the genotype - through a genotype initializer
  eoRealInitBounded<EOT>& init = make_genotype(parser, state, EOT());
  // Build the variation operator (any seq/prop construct)
  eoGenOp<EOT>& op = make_op(parser, state, init);
  // initialize the population - and evaluate
  // yes, this is representation indepedent once you have an eoInit
  eoPop<EOT>& pop   = make_pop(parser, state, init);
  // stopping criteria
  eoContinue<EOT> & term = make_continue(parser, state, eval);
  // output
  eoCheckPoint<EOT> & checkpoint = make_checkpoint(parser, state, eval, term);
  // algorithm (need the operator!)
  eoAlgo<EOT>& ea = make_algo_scalar(parser, state, eval, checkpoint, op);
  make_help(parser); //print help, if something is missing or user gives /h 
  // evaluate intial population AFTER help and status in case it takes time
  apply<EOT>(eval, pop);
  // print it out
  cout << "Initial Population\n";
  pop.sortedPrintOn(cout);
  cout << endl;

  run_ea(ea, pop); // run the ea

  cout << "Final Population\n";
  pop.sortedPrintOn(cout);
  cout << endl;

You can get away without understanding a single line of code above!! The only critical part is the following function, specified in the header file real_value.h. In order to adopt these versatile algorithms to solve a problem of our choosing, only changing this header file is adequate.

Fitness function

double real_value(const std::vector<double>& _ind)
{
  double sum = 0;
  for (unsigned i = 0; i < _ind.size(); i++)
      sum += _ind[i] * _ind[i];
  return sqrt(sum);
}

The eo library expects this function to be present and uses to evaluate the individuals in the population for fitness.

What happens here?

1. Our genotype vector has (say) n number of items. The fitness function simply squares each of these and add them up and returns the square root. No rocket science here! Just, the rule here is larger the individuals -- better the fitness!!
2. The GA code does the rest,

Try running the algorithm. It will go on decreasing the cost function and exit at 100th generation. The default is 100 generations, in a moment, we'll get into the details on how to change the default behavior.

Providing arguments

Once you run the above code once successfully, check the folder that consists the executable file.