Traditional evolutionary algorithms (EAs) treat every optimization problem as an isolated island. Whether designing a neural network or scheduling a logistics fleet, the algorithm starts from scratch, wandering through a high-dimensional landscape until it stumbles upon a peak. If the problem changes slightly, the process repeats, discarding every lesson learned from the previous search.
Evolutionary Multitasking (EMT), a paradigm championed by researchers like Abhishek Gupta and Yew Soon Ong, breaks this isolation. It allows a single population of agents to solve multiple related tasks simultaneously by sharing "genetic" knowledge. The intuition is simple: if two problems share structural similarities, the search trajectory of one can provide a shortcut for the other.
The Unified Search Space
The core mechanism of EMT is the creation of a unified search space. Instead of maintaining separate populations for Task A and Task B, the algorithm maps both problems into a shared coordinate system. In this environment, individuals are assigned a "skill factor" that determines which task they are currently optimizing.
As described in Evolutionary Multi-Task Optimization, this setup enables assortative mating. When two individuals from different tasks undergo crossover, they exchange genetic material that may contain "building blocks" useful to both. For example, a routing strategy that works well for a small city might contain a sub-sequence of stops that is also optimal for a larger metropolitan area. By solving these tasks concurrently, the algorithm exploits these shared landscape features to accelerate convergence.
Stepping Stones and Auxiliary Proxies
One of the most powerful applications of this idea is the use of auxiliary tasks. Often, a primary optimization goal is too complex or "deceptive" for an algorithm to solve directly. In these cases, researchers like Li Guo and Chao Wang 0099 suggest solving a simplified version of the problem alongside the main one.
In Evolutionary Multitasking AUC Optimization, the algorithm optimizes a "cheap" proxy (a small subset of data) and an "expensive" primary task (the full dataset) at the same time. The proxy task acts as a scout, rapidly identifying high-performance regions of the search space. These insights are then transferred to the primary task, helping it avoid local optima and reach global solutions with up to 90% fewer computational resources.
This concept of "stepping stones" is also central to the POET framework by Joel Lehman and Jeff Clune. In POET, agents are periodically transferred between diverse environments. A solution that stalls in a difficult environment might find a breakthrough in a simpler one, which then serves as a superior starting point when transferred back to the original challenge.
Mechanisms of Transfer
How exactly does "knowledge" move between tasks? The research identifies several distinct modalities:
- Genetic Crossover: In unified spaces, crossover naturally facilitates the exchange of traits. Abhishek Gupta shows that this implicit parallelism allows populations to "skill up" across tasks without explicit instructions.
- Explicit Migration: In frameworks like TLTL developed by Xiaoliang Ma, elite individuals are directly injected from one task's population into another. This structured transfer uses proven solutions as high-quality starting points rather than relying on random genetic luck.
- Iterative Process Transfer: Peng Li argues that we shouldn't just transfer final solutions, but the process of improvement. By sharing the sequence of transformations that led to a better result, the algorithm captures the underlying logic of the search space.
- Distributional Transfer: For black-box optimization, Yanchi Li and Wenyin Gong propose transferring the statistical "shape" of the search (mean and covariance matrices). This allows a task to inherit the search direction and step size of a more successful neighbor.
Avoiding the Multi-Task Curse
A major risk in multitasking is negative transfer, where information from a dissimilar task "pollutes" the search and leads to performance collapse. This is often called the "Multi-Task Curse."
To combat this, modern frameworks like MES-RET Yanchi Li use "safety valves" and adaptive mechanisms. These systems measure the "distance" or correlation between tasks in real-time. If two tasks are found to be unrelated, the transfer is throttled. Conversely, if tasks are highly synergistic, the algorithm increases the volume of shared data, ensuring that the multitasking overhead always results in a net gain in efficiency.
By treating optimization not as a series of isolated events but as a continuous process of learning and sharing, evolutionary multitasking transforms how we approach complex, high-dimensional problems—from Neural Architecture Search to Large-Scale Virtual Machine Placement.