Basic Optimization Framework - FORTRAN 90

Small project for relearning FORTRAN 90 and studying optimization concepts. This project is a generic framework over which simple optimization methods can be built upon and run.

NOTE: this project is finished.

Code structure:

The way this code is structured: - Optimization methods are implemented as subroutines, defined in the "contains" section of module optmod. Their names are limited to 20 characters (method_names variable). - Objective function definition, as well as its derivatives (when necessary) must be defined in file funcmod.f90, in the funcmod module. - The desired optimization method to be used, as well as its corresponding hyper parameters (see each method's definition) must be declared in the main.txt file. According to the following formatting:

   > Name of the optimization method (name of the subroutine)
   > Name of the hyper parameter 1 [tab or space] Hyper
     parameter 1 value
   > Name of the hyper parameter 2 [tab or space] 
     Hyper parameter 2 value
   > ...
   > Name of the next optimization method (and so on...)

• Array hyper parameters must have a space or tab separating each element, and the whole array must be contained in one line. Example:

   GD              [Gradient Descent method] 
   x   1   0   2   [Initial guess hyper parameter]
   LR  1D-3        [Learning Rate hyper parameter]
   eps 1D-10       [Relative error hyper parameter]
   freq 3          [How many this optim. should be repeated]
   GD              [Another optimization, also with the GD method]
   x   0   0   0   [Different initial guess]
   LR  1D-3                
   eps 1D-10               
   freq 2                  [Different number of repetitions]

How the code works

The mentioned procedures are all defined in the optmod module.

• Subroutine handle_main_file(...) handles reading of this file, calling the optimization methods and passing the important hyperparameters to them. As of know, they are written to a text file called output/hyperparams.txt (deleted by the end of execution), which is then read by the corresponding optimization method - this could be eventually optimized as described in the handle_main_file(...) subroutine documentation.

• Subroutine handle_hyp_file(...) handles the reading of the output/hyperparams.txt text file, attributing the correct values to the correct lines in an array. It is generic and called by each optimization method, which in turn interprets its results.

• Subroutine check_status(...) is only defined because it is commonly used code, and serves to handle file reading errors.

• Subroutine check_hprparams(...) checks if all required hyper parameters were informed by the user, and halts the program if not. It also warns the user if optional hyper parameters were not set, informing their default value.

NOTE: Default values for optional hyper parameters must be defined in the optimization procedure before calling the handle_hyp_file(...) subroutine, by initializing the associated elements of the hprdata array.

• It is best if optional hyper parameters are defined last (in hprnames), but this isn't a requirement. It is just an organization guideline.

• The arrays passed to handle_hyp_file(...) are defined as allocatable in the optimization methods because they depend on the dimension of the objective function's input --- dimx --- this because one of the hyper parameters is always the initial guess. It could be given special treatment, but this approach generalizes the process and may be useful in the future, in case multidimensional hyper parameters are used in future methods.

How to add to it

ADDING NEW METHODS: Other than updating optmod variables relating to the number of methods and their names (nb_methods and method_names), one must then update handle_main_file(...) subroutine's select case(optim_method) with the name of the new method. The method should also support the existing report parameters (check rprt_params) in the main program.

TODO: Add a reporting subroutine