Summary

FeiLiu36/EoH addresses the challenge of manual heuristic design by automating the process using a novel platform that synergizes Large Language Models (LLMs) with Evolutionary Computation (EC). It enables researchers and practitioners to efficiently design highly competitive algorithms and heuristics for complex search and optimization problems, significantly reducing development time and effort.

How It Works

EoH employs a coevolutionary framework where LLMs generate potential heuristic "thoughts" and corresponding code snippets, which are then evolved and refined by EC algorithms. This approach allows for the discovery of novel and effective heuristics by exploring a vast design space, outperforming traditional methods and even other automated approaches like FunSearch in terms of efficiency and solution quality.

Quick Start & Requirements

Installation involves cloning the repository and running pip install . within a Python >= 3.10 environment (conda recommended). Key dependencies include numba, numpy, and joblib. A crucial setup step is configuring access to either remote LLM APIs (e.g., OpenAI, Deepseek) or a locally deployed LLM. The project provides example usage for problems like TSP and online bin packing. Official documentation is available via the project's wiki.

Highlighted Details

EoH has demonstrated state-of-the-art performance, setting a new world record in the Circle Packing Problem and achieving an ICML 2024 Oral presentation (Top 1.5%). It significantly surpasses FunSearch in efficiency for the online bin packing problem, requiring fewer LLM queries. The project also has research accepted at PPSN 2024.

Maintenance & Community

The repository's maintenance frequency is noted as low, with a recommendation to explore the successor platform, LLM4AD, for more general-purpose applications. Community engagement links such as Discord or Slack are not explicitly provided in the README.

Licensing & Compatibility

EoH is released under the MIT License, which is permissive and generally suitable for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

The primary caveat is the low maintenance frequency, indicating potential for slower updates or bug fixes. Furthermore, the system's functionality is heavily dependent on the availability and configuration of LLM services, which can introduce external costs or complex setup requirements. The project also points towards LLM4AD as a more comprehensive future platform.