Evolutionary algorithms (EAs) and swarm intelligence metaheuristics are powerful tools for navigating complex, non-convex optimization landscapes. However, their success often hinges on a set of "magic numbers"—hyperparameters such as mutation rates, population sizes, step sizes, and crossover probabilities.
As Giovanni Iacca et al. note in Reinforcement learning based adaptive metaheuristics (GECCO Companion 2022), finding these configurations is traditionally a "dark art." Practitioners usually rely on manual trial-and-error to find a static set of parameters before running the algorithm. But this static approach has a fundamental flaw: the optimal parameter configuration is a moving target.
The Flaw of the Fixed Hyperparameter
To understand why static parameters fail, consider the lifecycle of an evolutionary search. In the early generations, the algorithm needs to prioritize exploration—taking large steps across the fitness landscape to discover promising regions and avoid getting trapped in local optima. In the later stages, the priority shifts to exploitation—narrowing the focus to fine-tune solutions near the global optimum.
A fixed parameter configuration forces a compromise. If the mutation rate is too high, the algorithm explores effectively but fails to converge; if it is too low, it converges prematurely. As Frank Neumann et al. point out in What's Hot in Evolutionary Computation (aaai 2017), fixed configurations cannot adapt to these shifting landscape requirements, leading to sub-optimal performance.
Furthermore, parameter sensitivity is highly instance-dependent. Jacques Bibaï et al. demonstrate in On the generality of parameter tuning in evolutionary planning (GECCO 2010) and Divide-And-Evolve Facing State-of-the-Art Temporal Planners during the 6th International Planning Competition (EvoCOP 2009) that tuning an algorithm for a broad problem domain fails to optimize for individual instances within that domain. This lack of generalizability creates a bottleneck when deploying metaheuristics autonomously across diverse real-world applications, from industrial disassembly planning to clinical image processing (Jin Xie et al., Disassembly sequence planning based on a modified grey wolf optimizer (Research 2021); Jorge Munilla et al., Detection of optic disc in human retinal images utilizing the Bitterling Fish Optimization (BFO) algorithm (Scientific 2024)). Even advanced frameworks like Natural Evolution Strategies struggle with the manual overhead of tuning population refresh and learning rates (Yi Sun et al., Efficient Natural Evolution Strategies (arxiv 2012)).
To bypass this bottleneck, a major research direction focuses on dynamic parameter control: mechanisms that allow algorithms to adjust their own hyperparameters on-the-fly.
Self-Adaptation and Feedback Loops
The most direct way to implement dynamic control is through feedback loops where the algorithm monitors its own progress and updates its parameters accordingly.
A classic example is the "one-fifth success rule," adapted for discrete optimization landscapes. In Optimal Parameter Choices Through Self-Adjustment: Applying the 1/5-th Rule in Discrete Settings (GECCO 2015), Benjamin Doerr et al. show that by measuring the success rate of recent mutations, an algorithm can dynamically scale its population size. If more than one-fifth of mutations are successful, the search is in an easy region and can expand exploration; if fewer are successful, it narrows down to exploit. This simple self-adjustment yields optimal linear runtime on benchmark problems, outperforming any static choice.
Another approach is to embed the hyperparameters directly into the evolutionary process. By appending the mutation rate to an individual's genome, the parameter itself becomes subject to selection pressure. In Runtime analysis for self-adaptive mutation rates (GECCO 2018), the authors prove that mutating the mutation rate allows the algorithm to automatically track the optimal rate throughout the run. This internal feedback loop provides exponential speedups over fixed rates, particularly in non-elitist populations where different phases of the search demand radically different strategies (Duc-Cuong Dang et al., Self-adaptation of Mutation Rates in Non-elitist Populations (arxiv 2016)).
The Surprising Power of Random Variation
Do these dynamic adjustments actually require an intelligent, tracking strategy? A counter-intuitive insight emerged when researchers investigated whether the benefits of parameter control stem from the strategy or simply from luck.
Giorgos Karafotias et al. demonstrated in Parameter control: strategy or luck? (GECCO (Companion) 2013) that merely varying parameters randomly throughout a run—without any feedback loop—can match or even outperform sophisticated control strategies. This suggests that simply breaking the rigidity of a static parameter is inherently beneficial.
Building on this idea, Denis Antipov et al. introduced a "lazy" parameter control method in Lazy parameter tuning and control: choosing all parameters randomly from a power-law distribution (GECCO 2021). Instead of tracking state transitions, the algorithm samples its hyperparameters from a heavy-tailed power-law distribution at each step.
Power-Law Distribution Sampling:
- High probability: Small parameter values (Local exploitation)
- Low probability: Massive parameter values (Global exploration leaps)
Because of the mathematical properties of power laws, the algorithm naturally spends most of its time executing small, exploitative steps, while occasionally taking massive, exploratory leaps. This stochastic variation completely bypasses the need for complex tracking mechanisms while remaining highly competitive across diverse problem instances.
Learning to Control via Reinforcement Learning
While random variation works well for single parameters, complex multi-objective landscapes often require coordinated control over multiple parameters simultaneously (e.g., balancing mutation, crossover, and selection pressure at once). Handcrafting separate rules for each parameter leads to an ad-hoc "patchwork problem," where control mechanisms conflict with one another (Giorgos Karafotias et al., Generic parameter control with reinforcement learning (GECCO 2014)).
To achieve unified control, researchers have turned to Reinforcement Learning (RL). By treating the evolutionary algorithm as the environment and the parameter controller as the agent, RL can learn an optimized control policy. The state space consists of population statistics (such as fitness diversity), and the actions are adjustments to the hyperparameters.
Modern iterations of this paradigm leverage deep learning to handle highly complex optimization spaces. For example, Robbert Reijnen et al. developed GS-MODAC in Graph-Supported Dynamic Algorithm Configuration for Multi-Objective Combinatorial Optimization (icml 2025).
[Evolving Population] ──> [Graph Neural Network] ──> [Deep RL Agent] ──> [Dynamic Hyperparameters]
By modeling the evolving population as a graph in the objective space, a Graph Neural Network (GNN) captures structural patterns of convergence. A Deep RL agent uses these patterns to adjust mutation and crossover rates in real-time, allowing the controller to generalize successfully to unseen problem sizes.
This intersection of metaheuristics and deep learning has even come full circle, with adaptive evolutionary controllers being used to optimize deep learning architectures themselves. In ACO-MoE-LoRA: Evolving-while-Training for Adapting Segment Anything Model 2 to Specialized Domains (icml 2026), Kaiyi Luo et al. utilize an "Evolving-while-Training" framework. Here, an Ant Colony Optimization controller dynamically tunes model topologies and routing weights during training, demonstrating how adaptive metaheuristics can prevent structural collapse in complex Mixture-of-Experts architectures.
By moving from rigid, handcrafted configurations to dynamic, self-adjusting, and learned control policies, this research direction transforms hyperparameters from brittle liabilities into adaptive assets.
Key Foundational Work
Un état des lieux de l'optimisation évolutionnaire et de ses implications en sciences pour l'ingénieur (2007) — contributed 1 concept elements to this direction.
Divide-And-Evolve Facing State-of-the-Art Temporal Planners during the 6th International Planning Competition (2009) — contributed 1 concept elements to this direction.