[Paper] Energy-Aware Metaheuristics

Source: arXiv

Source: arXiv:2602.06595v1

Overview

The paper introduces a systematic way to make metaheuristic algorithms energy‑aware, meaning they can operate under a strict power or energy budget while still delivering high‑quality solutions. By modeling each algorithmic operator in terms of both its expected improvement and its energy consumption, the authors enable solvers to automatically pick the most “bang‑for‑your‑Joule” actions during the search.

This is especially relevant as AI‑driven optimisation moves onto edge devices, IoT sensors, and other energy‑constrained platforms.

Key Contributions

• Unified operator‑level model – simultaneously quantifies numerical gain (expected improvement) and energy usage for any metaheuristic operator.
• Expected Improvement per Joule (EI/J) metric – drives adaptive operator selection, balancing exploration vs. exploitation under a fixed energy budget.
• Energy‑aware extensions of three classic metaheuristics:
  • Steady‑state Genetic Algorithm
  • Particle Swarm Optimisation
  • Iterated Local Search
    each equipped with lightweight and heavyweight operator variants.
• Extensive empirical validation on three diverse combinatorial problems:
  • 0/1 Knapsack
  • NK‑landscapes
  • Error‑Correcting Codes
    Results show comparable solution quality with significantly lower energy consumption.
• Insightful analysis of EI/J dynamics – demonstrates early stabilization and clear patterns that enable each solver to autonomously discover the most efficient operator for a given problem.

Methodology

1. Operator Profiling – For each candidate operator (e.g., mutation, crossover, velocity update) the authors measure two quantities:

   • Expected Improvement (EI) – the average increase in fitness observed when the operator is applied.
   • Energy Cost – the Joules consumed per application (measured on a reference hardware platform).

2. EI/J Score – Compute the ratio

   EI / EnergyCost

   This yields a scalar “efficiency” score. Higher values indicate more improvement per unit of energy.

3. Adaptive Selection Loop – During the optimisation run, the algorithm keeps a small sliding window of recent EI/J estimates for each operator. At every decision point it selects the operator with the highest current EI/J, allowing the search to self‑regulate between cheap, low‑gain moves and expensive, high‑gain moves.

4. Energy‑Budget Enforcement – The total energy available is fixed at the start of the run. The algorithm terminates once the cumulative energy consumption reaches this budget, guaranteeing that the search never exceeds the allocated power envelope.

5. Instantiation – The framework is integrated into three metaheuristics:

   Metaheuristic	Operators Compared
   Steady‑state GA	Light mutation vs. heavy crossover
   Particle Swarm Optimisation (PSO)	Cheap velocity update vs. costly, globally‑aware update
   Iterated Local Search (ILS)	Fast local perturbation vs. thorough, computationally intensive neighbourhood search

The methodology thus provides a principled way to balance optimisation performance against energy consumption, enabling energy‑aware metaheuristic search.

Results & Findings

• EI/J stabilisation occurs after only a few dozen iterations; thereafter the algorithm consistently favours the most efficient operator.
• The operator‑selection pattern is problem‑dependent:
  • On the Knapsack problem the heavy crossover quickly dominates.
  • On NK‑landscapes the lightweight mutation remains preferred for most of the run.
• Across all three metaheuristics, the energy‑aware versions achieve comparable solution quality while consuming roughly one‑third less energy than their standard counterparts.

Practical Implications

• Edge AI & Embedded Optimisation – Devices such as drones, wearables, or smart sensors can now run sophisticated combinatorial optimisation (routing, scheduling, resource allocation) without draining batteries prematurely.

• Green Computing – Data‑center workloads that involve large‑scale metaheuristic searches (e.g., hyper‑parameter tuning, layout optimisation) can be throttled to a predefined power envelope, reducing operational carbon footprints.

• Adaptive Power‑Management APIs – The EI/J metric can be exposed as a plug‑in for existing optimisation libraries (DEAP, PyGMO, Optuna), allowing developers to toggle “energy‑aware mode” with a single flag.

• Self‑Optimising Systems – Because the algorithm learns on‑the‑fly which operators are most cost‑effective, it can adapt to hardware heterogeneity (CPU vs. GPU vs. ASIC) without manual retuning.

Limitations & Future Work

• Hardware Dependence – Energy measurements were taken on a specific benchmark platform; transferring EI/J scores to vastly different architectures (e.g., ARM microcontrollers vs. GPUs) may require recalibration.
• Operator Granularity – The framework assumes a discrete set of pre‑defined operator variants; extending it to continuous parameter spaces (e.g., mutation‑rate tuning) remains an open challenge.
• Scalability to Massive Budgets – While the method shines under tight budgets, its benefits diminish when energy is abundant, as the baseline already reaches near‑optimal fitness.

Future Directions (as suggested by the authors)

• Integrating online energy prediction models.
• Exploring multi‑objective extensions (e.g., balancing time, energy, and solution quality).
• Applying the approach to real‑time control problems, such as autonomous‑vehicle path planning.

## Authors

- **Enrique Alba**
- **Tomohiro Harada**
- **Gabriel Luque**

Paper Information