Evolutionary Manytasking Optimization Based on Symbiosis in Biocoenosis

Rung-Tzuo Liaw, Chuan-Kang TingAAAI 2019 · DOI

Abstract

This paper introduces Symbiosis in Biocoenosis Optimization (SBO), a bio-inspired framework designed for evolutionary many-tasking. By treating multiple optimization tasks as a community of interacting species, SBO adaptively manages information transfer through biological relationship models like mutualism and parasitism. Experimental results on 30-task benchmarks demonstrate that SBO achieves superior convergence and solution quality compared to state-of-the-art multitasking algorithms.

Introduction

Evolutionary Algorithms (EAs) have long been a staple for complex optimization. Traditionally, these algorithms focus on solving a single problem at a time. The field of Evolutionary Multitasking (EMT) emerged to address multiple related problems concurrently, leveraging information sharing to accelerate search. However, current state-of-the-art methods, such as the Multi-factorial Evolutionary Algorithm (MFEA), struggle with scalability when moving from a few tasks to "many-tasking" scenarios involving 30 or more tasks. These methods often suffer from rigid information-sharing mechanisms and algorithmic constraints that limit their applicability to diverse EA types.

The SBO Framework

The authors propose Symbiosis in Biocoenosis Optimization (SBO), a framework that conceptualizes a set of optimization tasks as a biocoenosis—an ecological community of different species. In this model, each species represents a population dedicated to a specific optimization task.

Conceptual illustration of the Biocoenosis framework where species (tasks) interact through Symbiosis.

The core of SBO is the adaptive management of inter-task interactions through Symbiosis:

  • Algorithm-Agnostic Transfer: Instead of relying on genetic crossover, SBO uses individual replacement. A portion of a task's offspring is replaced by high-performing individuals from another task, making the framework compatible with any EA, including those without recombination operators.
  • Symbiosis Estimation: The relationship between tasks is dynamically categorized into biological types: Mutualism, Commensalism, Parasitism, Neutralism, Amensalism, and Competition. This is achieved through "paired evaluations" to determine if transferred individuals benefit or hinder the target task.
  • Adaptive Control: The transfer rate is adjusted based on these estimations. Beneficial relationships (Mutualism) lead to increased transfer, while harmful ones (Parasitism) are throttled to prevent negative transfer.

Experimental Evaluation

The SBO framework was tested using a suite of Many-tasking Problems (MaTPs) consisting of 30 benchmark functions from the CEC 2017 competition. To evaluate robustness, four variations of the suite ($MaTP_Z, MaTP_S, MaTP_M, MaTP_L$) were created, representing different degrees of shift between the optimal solutions of the tasks.

Results and Discussion

SBO was compared against MFEA and several other baseline algorithms using both Genetic Algorithms (GA) and CMA-ES as base solvers.

Convergence curves showing Mean Best Fitness across different task suites (MaTP_Z, MaTP_S, MaTP_M, MaTP_L) for selected functions.

As shown in the convergence plots, SBO consistently outperformed competing methods across nearly all 30-task benchmarks. In the $MaTP_Z$ suite, which features high task similarity, SBO achieved superior results in 29 or 30 out of 30 tasks. Key findings include:

  • Superior Solution Quality: SBO reached significantly better final fitness values than MFEA.
  • Faster Convergence: The adaptive symbiosis mechanism allowed SBO to find high-quality solutions much earlier in the evolutionary process.
  • Effective Filtering: Analysis of survival rates confirmed that SBO successfully filters out "noisy" transfers, ensuring that only useful genetic material is shared between tasks.

Conclusion

SBO represents a significant advancement in evolutionary many-tasking by treating optimization as a dynamic ecosystem. By replacing rigid transfer rules with adaptive biological models, it effectively mitigates the risk of negative transfer while scaling to dozens of tasks. Future research may explore heterogeneous biocoenosis, where different species utilize different underlying optimization strategies to further enhance performance on diverse real-world problems.

Summary updated

BibTeX
@inproceedings{liaw2019evolutionary,
  author = {Rung-Tzuo Liaw and Chuan-Kang Ting},
  title = {{Evolutionary Manytasking Optimization Based on Symbiosis in Biocoenosis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33014295},
}
Figure