Architecture

EvoAl core implementation consists of multiple modules (realised as Maven projects) that provide different APIs and functionality. In the following, we will sketch the currently existing modules, their purpose, and their dependencies (among each other).

flowchart LR
    core.main
    core.junit
    core.arff
    generator.main
    optimisation.api
    optimisation.ea
    optimisation.pso
    surrogate.api
    surrogate.simple
    surrogate.smile
    surrogate.adaption

    core.arff --> core.main
    core.junit --> core.main

    generator.main --> core.main
    generator.main --> optimisation.api

    optimisation.api --> core.main
    optimisation.ea --> optimisation.api
    optimisation.pso --> optimisation.api

    surrogate.api --> core.main
    surrogate.api --> optimisation.api
    surrogate.simple --> surrogate.api
    surrogate.smile --> surrogate.api
    surrogate.adaption --> surrogate.api
    surrogate.adaption --> optimisation.api

core.main

The module core.main contains the most fundamental concepts of EvoAl, such as the Blackboard, the CDI orchestration, model handling, and DSL handling. This module contains no optimisation or machine-learning-specific code to keep EvoAl as flexible as possible.

core.junit

The module core.junit contains some helpers for unit testing, such as resource loading, DSL configuration, and DSL loading from the src/test/resources folder (for regression testing).

core.jacoco

The module core.jacoco (not mentioned in the overview) is a technical artifact that is necessary to do proper code coverage measurement.

core.arff

The module core.arff contains code to support the Weka ARFF format for datasets.

generator.main

The module generator.main implements a data generation framework based on benchmark functions typically used in the optimisation domain. Therefore, it uses the Generator DSL. Additionally, the benchmark functions can be used as fitness functions in the optimisation use-case.

optimisation.api

The module optimisation.api provides the core concepts that are common to all optimisation algorithms, such as fitness functions, initial population providers, and so on. Any optimisation algorithm uses these elements to be able to reuse existing implementations and extensions, such as benchmark functions or surrogate models.

optimisation.ea

The module optimisation.ea implements the evolutionary algorithm and its specialisations, such as, genetic programming.

optimisation.pso

The particle swam optimisation algorithm is implemented in the module optimisation.pso.

surrogate.api

API and default implementations for surrogate learning (and usage) can be found in the surrogate.api module. The module parses the training specification, loads data, starts the training process, and calls quality of fit calculators.

surrogate.simple

The module surrogate.simple implements some very simple model learners.

surrogate.smile

The module surrogate.smile uses the SMILE library to offer support vector regression.

surrogate.adaption

Surrogate adaption techniques, based on the density of training data or the prediction interval are implemented in the module surrogate.adaption.