Evolving Objects: A General Purpose Evolutionary Computation Library

Maarten Keijzer, Juan Julián Merelo Guervós, Gustavo Romero, Marc SchoenauerArtificial Evolution 2001 · DOI

TL;DR

This paper introduces EOlib, an object-oriented C++ library designed to make evolutionary computation (EC) more flexible and accessible. EOlib allows researchers to evolve virtually any data structure, not just predefined ones, by focusing on interfaces and enabling the creation of custom representations and variation operators. This aims to overcome the limitations of existing EC libraries and foster reproducibility and innovation in the field.

What Problem Does This Paper Address?

Evolutionary Algorithms (EAs) are powerful optimization techniques inspired by natural evolution. However, applying them to real-world problems often requires representing complex data structures (like neural networks or fuzzy logic controllers) in a way that existing EC libraries can handle. Most libraries are restrictive, forcing users to "flatten" their data into standard formats (binary strings, arrays) or limiting the types of variation operators (crossover, mutation) that can be used. This inflexibility leads to a dilemma: either comply with the library's limitations or build a custom solution from scratch, which is time-consuming and prone to introducing new, unforeseen limitations. The authors aim to bridge this gap by providing a flexible, paradigm-free framework that allows for the evolution of arbitrary objects.

Methods

The core of EOlib's approach is its object-oriented design, treating every component of an EC system – data structures, operators, and even statistics – as an object. This "representation-independent" and "paradigm-independent" philosophy is central to its flexibility.

Key concepts include:

  • Evolvable Objects: Any object can be evolved if it possesses certain properties:
  • Initialisability: The ability to create new objects, often randomly.
  • Selectability: The object must have a fitness value that can be used for selection.
  • Replicability: The ability to create copies of the object.
  • Mutability/Combinability: The object must be able to change (mutation) or be combined with others (crossover/recombination).
  • Interfaces: EOlib focuses on defining interfaces that objects must implement to be evolvable. This allows for a wide range of data structures without needing to know their internal details.
  • Object-Oriented Components:
    • Variation Operators: These are separate objects that act on evolvable objects. They can take any number of inputs and produce any number of outputs, allowing for complex combinations and custom operators.
  • Evolution Engines: These manage the population lifecycle (initialization, selection, variation, replacement). EOlib supports various standard engines but also allows users to define new ones by assembling existing building blocks.
  • Fitness Functions: EOlib supports more than just scalar fitness, allowing for multi-objective optimization and constraint handling.

Datasets

Not discussed in the paper. The paper focuses on the framework and its applications rather than specific benchmark datasets.

Key Insights and Ideas

  • "Evolve Anything" Philosophy: The most significant idea is the shift from evolving predefined representations to evolving the actual data structures of interest. This dramatically broadens the applicability of EC.
  • Interface-Driven Design: By defining clear interfaces for evolvability, EOlib decouples the EC algorithm from the specific representation, promoting reusability and extensibility.
  • EC as Objects: Treating every aspect of an EC system as an object allows for unprecedented modularity and flexibility. This includes the ability to evolve the EC algorithms themselves (multilevel EC).
  • Reproducibility: The paper highlights the potential for EOlib to improve the reproducibility of EC research by making source code and experimental setups readily available.
  • Beyond Standard EC: EOlib is designed to go beyond traditional Genetic Algorithms, Evolution Strategies, etc., by supporting novel variation operators, fitness functions, and population management strategies.

Key Results

The paper showcases EOlib's flexibility through several applications:

  • Evolving Neural Networks: Directly evolving multilayer perceptrons without flattening them into arrays.
  • Genetic Mastermind: A GA was programmed to find hidden combinations, achieving solutions faster than previous implementations.
  • Fuzzy Logic Controllers: Evolving bidimensional fuzzy logic controllers to approximate functions.
  • RBF Neural Networks: Evolving RBF networks with diverse radii, resulting in smaller and more accurate models.
  • Image Segmentation: Applying EC to fruit fly embryo image processing.
  • Topological Optimum Design: Implementing multi-objective optimization for structural design.

While specific quantitative results (e.g., percentage improvements) are mentioned for some applications, the primary focus is on demonstrating the feasibility and flexibility of the EOlib framework across diverse problem domains.

Limitations and Future Work

  • Completeness: The current version (0.9.1) is noted as not yet complete, indicated by the leading zero in the version number.
  • Future Work: The authors plan to:
  • Implement EOlib in other object-oriented languages like Java and explore interoperability.
  • Develop an application generator that uses high-level EC languages and an operating-system independent GUI.
  • Provide an object repository for users to share their EOlib-based components, further enhancing reproducibility and community contribution.

Summary updated

BibTeX
@inproceedings{keijzer2001evolving,
  author = {Maarten Keijzer and Juan Julián Merelo Guervós and Gustavo Romero and Marc Schoenauer},
  title = {{Evolving Objects: A General Purpose Evolutionary Computation Library}},
  booktitle = {Artificial Evolution},
  year = {2001},
  doi = {10.1007/3-540-46033-0\_19},
}
Figure