developer_guide.md

Developer's Guide

Creating new algorithms

Currently, you can extend evogil by adding new algorithms to evogil/algorithms package. In the future, there will be also a possibility to run external, custom algorithms' scripts, without integrating them into the framework code.

1. Create algorithm class by extending Driver:

   from algorithms.base.driver import Driver
   
   class MyAlgorithm(Driver):
       pass

2. Implement missing methods from Driver class:

   • step() - single step of algorithm. Steps are invoked by simulation runner. Typically step represents a single generation of an algorithm. However, if an algorithm hasn't generational nature (e.g. steady-state), it may be used as other repeatable chunk of calculations. Note that algorithms initialization should be invoked in constructor.
   • finalized_population() - this method is called by simulation runner at the end of simulation to obtain final state of algorithm's population. Population should be represented as list of individuals (vectors of real numbers with size equal to dimensions count of a problem domain).

3. Add a constructor with all parameters required by your algorithm. Parameter list has to be extended with *args, **kwargs params passed to super() (Driver) constructor. For example:

       def __init__(
           self,
           population,
           dims,
           fitnesses,
           mating_population_size,
           mutation_eta,
           crossover_eta,
           mutation_rate,
           crossover_rate,
           *args,
           **kwargs
       ):
           super().__init__(*args, **kwargs)
           # Further initialization...

4. Add your algorithm configuration to run_config.py script:

   • add the algorithm name to drivers list
   • add any specific to your algorithm parameters in algo_base dictionary
   • optionally, add custom configuration method with a name init_alg___X where X is your algorithm name. It is useful to fill typical parameters, such as mutation/crossover rate using default values, for example:

     def init_alg___MyAlgorithm(algo_config, problem_mod):
        standard_variance(algo_config, problem_mod)

5. Check if everything is ok simply by running evogil with your algorithm, for example:

   run budget 500 -a MyAlgorithm -p zdt1