Inter Task Knowledge Transfer in Evolutionary Optimization

98 concepts59 papers1 proposal

Traditional evolutionary algorithms treat optimization problems as isolated islands. Whether optimizing a neural network architecture, a vehicle routing schedule, or a high-dimensional feature subset, the algorithm typically starts from scratch, blindly exploring a complex fitness landscape.

However, real-world optimization problems rarely exist in a vacuum; they frequently share underlying structural topographies. As highlighted by Hangqi Ding et al. in their comprehensive survey, Evolutionary Multitasking (EMT) upends the traditional single-task paradigm. By solving multiple distinct optimization tasks concurrently within a single unified environment, EMT allows populations to actively share search experiences, providing a powerful mechanism to escape local optima and accelerate convergence.

The Intuition: Why Solve Multiple Problems at Once?

Why does adding more problems to an evolutionary run make it faster instead of harder? The core intuition lies in cross-task shortcuts.

When an evolutionary algorithm navigates a complex, high-dimensional landscape, it can easily stall in deceptive local optima. However, if a related auxiliary task is being solved simultaneously and has already mapped out a similar region of its own search space, its high-performing solutions can serve as a guide. By transferring this structural information, the struggling task receives a conceptual stepping stone, bypassing thousands of generations of redundant exploration.

Traditional Optimization:
Task A Population  --> [ Blind Exploration ] --> Local Optima Stagnation
Task B Population  --> [ Blind Exploration ] --> Slow Convergence

Evolutionary Multitasking:
Task A Population  ---\
                       ===> [ Inter-Task Knowledge Transfer ] ===> Accelerated Global Optima
Task B Population  ---/

This synergy is particularly effective when tasks exhibit complementary characteristics. As demonstrated by Eneko Osaba et al. in graph community detection, maintaining distinct subpopulations that periodically exchange individuals allows the optimization process to exploit overlapping structural similarities that would remain invisible to isolated searches.

Mechanisms of Knowledge Transfer

Knowledge cannot simply be dumped from one task into another; it must be transferred in a way that respects the unique geometry of each problem. The literature reveals two primary methodologies for executing this transfer: explicit solution crossover and implicit distribution alignment.

Explicit Solution Transfer (Genotypic Crossover)

The most direct way to share information is to allow individuals assigned to different tasks to mate. Chen Wang et al. utilize assortative mating, a mechanism where individuals are assigned "skill factors" for specific tasks but are stochastically permitted to undergo cross-task crossover. This allows beneficial structural patterns to propagate across task boundaries within a single population.

Similarly, Abhishek Gupta et al. show that transferring genetic material via crossover allows high-performing solutions from one task to serve as effective search heuristics for another. To make this process more efficient, Zhiwei Xu et al. introduce "Transfer Sparks," which calculate explicit direction vectors between the elite individuals of different tasks, providing a direct spatial shortcut across related search spaces.

Implicit Strategy and Distribution Transfer

Swapping raw individuals can sometimes backfire. As Zhengping Liang et al. point out, transferring specific individuals risks propagating solutions that are only locally optimal for a single task, which can mislead the recipient population.

To mitigate this, modern frameworks often transfer search space characteristics rather than explicit candidates:

  • Distribution Parameters: Dejun Xu et al. transfer the mean and covariance matrices of an evolutionary strategy between spatially close tasks, initializing new searches with informed topographies.
  • Evolutionary Hyperparameters: Sheng-Hao Wu et al. propose transferring successful evolution parameters (such as mutation and crossover rates) to help struggling tasks adopt search behaviors optimized for similar landscape types.
  • Input Warping: In expensive black-box scenarios, Jasper Snoek et al. demonstrate that learned input transformations can be transferred between low-fidelity proxies and high-fidelity targets to capture parameter sensitivity.

Manufacturing Relatedness via Problem Reformulation

One of the most powerful realizations in this field is that we do not need to wait for naturally related problems to appear; we can intentionally break a single complex problem down into a multitasking environment.

Decomposing Bi-Level and Multi-Echelon Problems

In bi-level optimization, lower-level tasks are nested inside upper-level decisions, creating an expensive computational bottleneck. Abhishek Gupta et al. treat these lower-level tasks as related sub-problems solved concurrently, slashing computational costs by over 50%. For supply chain logistics, Xueming Yan et al. decompose multi-echelon routing problems into distinct sub-tasks per echelon, using cross-task crossover to exploit inter-echelon correlations.

Managing Scale and Objectives

When dealing with hundreds of variables or conflicting goals, the search space becomes unmanageable. Handing Wang et al. introduce a framework that reformulates a single large-scale many-objective problem into multiple related sub-tasks, accelerating convergence by exploiting similarities across different parts of the objective space.

Furthermore, creating a simplified or proxy version of the main goal provides a necessary shortcut. Yaqing Hou et al. describe how solving auxiliary, simplified versions of a primary task simultaneously provides structural building blocks that prevent the algorithm from treating every new challenge as a blank slate.

Managing the Dynamics of Transfer

To prevent "negative transfer"—where an unrelated task misleads the search—algorithms must dynamically manage how and when information flows.

Chao Wang et al. implement adaptive knowledge transfer that migrates candidate solutions based on calculated relationship scores between tasks. Within individual solutions, Xiaoliang Ma et al. supplement inter-task transfer with intra-task transfer, sharing information between different decision variables within a single solution to exploit structural similarities across dimensions.

Finally, when optimization occurs in sequential phases, Huiting Li et al. emphasize the use of heterogeneous transfer learning to prevent information loss when transitioning between distinct optimization stages.

Applications Across Domains

The versatility of this paradigm is demonstrated across diverse machine learning and engineering fields:

  • Neural Architecture Search (NAS): Peng Liao et al. partition populations to continuously transfer structural blocks across different datasets, leveraging architectural insights from one image domain to improve performance on another.
  • High-Dimensional Feature Selection: Xubin Wang et al. split a particle swarm into cooperative sub-populations to share knowledge about feature importance, providing a shortcut through massive search spaces.
  • Physics-Informed Neural Networks (PINNs): Bo Wang et al. apply multi-task optimization to train PINNs, using learnable coefficients to adaptively combine parameters from auxiliary tasks, which significantly improves generalization when data is scarce.

Key Foundational Work

Learn-and-Optimize: A Parameter Tuning Framework for Evolutionary AI Planning (2011) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 59 papers
  2. Version 1 · 2026-10-04 · 62 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_evolutionary_multitasking_b48ecefc,
  title = {evolutionary multitasking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/evolutionary-multitasking/rdr_b48ecefc0fed408690cca11f1fd2bc06},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_evolutionary_multitasking_b48ecefc_v2,
  title = {evolutionary multitasking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/evolutionary-multitasking/rdr_b48ecefc0fed408690cca11f1fd2bc06/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake