Motivation

Most optimization algorithms search for a single best solution, but this can be suboptimal for various reasons. First, many optimizers get stuck in a local min/maximum searching for a single best solution. Second, a lot of optimizers are unable to balance exploration and exploitation, which can lead them to overlook promising solutions or even get stuck in local optima. Third, almost all optimization algorithms are not ammenable to understanding the relationship between features and performance — they are bascially just black boxes. And finally, and arguably the biggest motivation, is that traditional optimization seeks a single best solution for a given situation. However, what works well in one situation might not be the best solution for another. Quality-diversity algorithms address these shortcomings by searching for a diverse set of high performing solutions, rather than a single solution. One of the most successful and widely used quality-diversity algorithms is Multi-dimensional Archive of Phenotypic Elites (MAP-Elites), which can be categorized as an evolutionary algorithm. MAP-Elites uses genetic operators to evolve high quality candidate solutions at each point within a discretized behavior (feature) space, which the user can choose.^[1]

Package

• Modularity: MAPElites.jl includes various types of archives, behavior descriptor functions, fitness functions, and variation operators that you can mix and match however you see fit. Furthermore, you can easily extend MAPElites.jl's functionality by implementing your own components and mixing and matching them with existing components.

• Simplicity at its core: While MAPElites.jl has many modular components for you to mix and matched, you really only need one struct to use them. Additionally, several MAP-Elites variants are implemented that you can run with on function call.

• Library-agnostic: Unlike some Python implementations of quality-diversity algorithms, MAPElites.jl doesn't require you to use any specific library for arrays or tensors. You are free to use base Julia, CUDA arrays from CUDA.jl, or any other library. As long as your fitness function can evaluate your candidate solutions, you can use whatever data structure you want, be it arrays, strings, images, or anything else.

• Works with Optimization.jl: The implementations in MAPElites.jl can also be called from within Optimization.jl (coming soon).

Alternatives

There are a few alternatives to MAPElites.jl for quality-diversity algorithms written in Python.

• QDax: A Jax-based implementation of several quality-diversity algorithms mostly designed around reinforcement learning and robotics.

• PyRibs: A barebones Python implementation that uses an ask-tell interface that also focuses on reinforcement learning.

• qdpy: A generic minimal implementation of quality-diversity algroithms in Python.

• CVT-MAP-Elites: A simple Python implementation of CVT-MAP-Elites.

Contents

• Changelog
• • [Unreleased]
• MAPElites.jl
• MAPElites.jl APIs
• • AbstractArchive
  • VariationStrategy
  • FitnessStrategy
  • BehaviorDescriptorStrategy
• Modules
• Abstract Types
• Structs
• Functions
• Contributing
• • Reporting a Bug
  • Requesting New Features
  • Contributing Code
  • Updating or Fixing Documentation
• Installation
• Usage
• • Standard Workflow
  • Mixing and Matching Components
  • Using with Optimizers.jl
• Algorithms
• • MAP-Elites
  • CVT-MAP-Elites
  • AURORA
  • SAIL
  • CMA-MAP-Elites
• References
• Motivation
• Package
• Alternatives
• Contents
• References

References