🤖 machine learning

QUASAR: An Evolutionary Algorithm to Accelerate High-Dimensional Numerical Optimization

This paper introduces QUASAR, a novel evolutionary algorithm that enhances Differential Evolution with quasi-adaptive stochastic mechanisms inspired by quantum particles to significantly outperform existing methods in solving high-dimensional, non-differentiable optimization problems.

Original authors: Julian G. Soltes

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Julian G. Soltes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the single best spot to set up a campfire in a massive, foggy forest. This forest is so big and complex that it has thousands of dimensions (like layers of hills, valleys, rivers, and trees stacked on top of each other). This is what scientists call "high-dimensional optimization."

If you try to walk in a straight line or follow a map (like traditional math methods), you'll get stuck or lost because the terrain is too weird. You need a team of explorers to search the whole forest at once. This is where QUASAR comes in.

Here is a simple breakdown of what the paper says QUASAR does, using everyday analogies:

The Problem: The "Curse of Dimensionality"

Think of the search space as a giant, multi-layered maze. As the maze gets more complex (more dimensions), it becomes exponentially harder to find the exit. Old methods often get stuck in a small dead-end (premature convergence) or wander aimlessly without ever finding the best spot.

The Solution: QUASAR (The Quantum-Inspired Explorer)

QUASAR is a new "team of explorers" (an evolutionary algorithm) designed to find the best solution faster and better than previous teams. It takes inspiration from how quantum particles behave—specifically, the idea that particles can exist in many places at once and "tunnel" through walls to get to better spots.

The paper claims QUASAR uses three main tricks to beat the competition (specifically two other famous algorithms called DE and L-SHADE):

1. The "Spooky" Teamwork (Mutation)

Imagine your team of explorers is split into three groups, and they decide who does what based on a coin flip (probability):

The "Spooky-Best" Group (Exploitation): These explorers stick close to the current best campsite found so far. They make tiny, careful adjustments to see if they can make the fire just a little bit warmer. They are the "perfectionists."
The "Spooky-Current" and "Spooky-Random" Groups (Exploration): These explorers are the adventurers. They jump far away from the current best spot to check out completely new, unexplored parts of the forest. They are the "risk-takers."
The Magic: The algorithm constantly switches between these groups. It doesn't just stick to one strategy; it dynamically balances being careful and being bold.

2. The "VIP" Mixing (Crossover)

When the explorers share their findings, the algorithm treats them differently based on how good their findings are.

If an explorer found a great spot, they get to mix their ideas with the new "Spooky" ideas very often.
If an explorer found a bad spot, they mix less.
The Analogy: Think of it like a recipe contest. The judges (the algorithm) let the chefs with the best dishes mix their secret ingredients with new ideas more often, while the chefs with bad dishes get fewer chances to change their recipes. This ensures the "best ideas" spread quickly.

3. The "Teleportation" Reset (Asymptotic Reinitialization)

This is the paper's biggest innovation. Imagine that 33% of your explorers are stuck in a swamp (the worst-performing solutions). Instead of just telling them to "try harder," QUASAR gives them a special teleportation device.

How it works: The algorithm looks at where the top 25% of explorers are standing. It creates a "map" (a covariance matrix) of that good area.
The Teleport: It instantly moves the stuck explorers to a random spot near the good area, rather than leaving them in the swamp.
The Twist: At the very beginning of the search, this teleportation happens very often because the explorers are lost. As the search goes on and the team gets smarter, the teleportation happens less often, letting the team focus on fine-tuning the best spot they found.

The Results: Did it work?

The author tested QUASAR against the other two famous algorithms (DE and L-SHADE) using a standard set of 29 very difficult math puzzles (the CEC2017 benchmark).

Better Quality: QUASAR found significantly better solutions. On average, the solutions it found were 3.85 times better than the standard DE and 2.07 times better than L-SHADE.
Faster Speed: QUASAR was also faster. It finished the tasks 1.4 times faster than DE and a massive 5.16 times faster than L-SHADE.
Easy to Use: The paper claims you don't need to be a math wizard to use it. It works well with its default settings, so you don't have to spend hours tweaking knobs and dials.

The Bottom Line

The paper concludes that QUASAR is a powerful, efficient, and easy-to-use tool for solving extremely complex problems where you have to find the "best" answer in a massive, confusing search space. It does this by mixing careful perfectionism, wild exploration, and a smart "teleportation" reset for the explorers who get stuck.

Note: The paper focuses entirely on mathematical optimization problems (like finding the best numbers for a formula). It does not claim to solve medical problems, predict the stock market, or cure diseases, though the author mentions these are fields where such math is used.

Technical Summary: QUASAR – An Evolutionary Algorithm for High-Dimensional Optimization

Problem Statement
High-dimensional numerical optimization remains a persistent challenge in computational science, particularly within fields such as machine learning, engineering, and finance. The primary obstacle is the "curse of dimensionality," where the exponential growth of the search space renders traditional gradient-based and parametric search methods ineffective. While metaheuristics, specifically Differential Evolution (DE), have established themselves as robust optimizers, standard DE variants often struggle to maintain an effective balance between global exploration and local exploitation in high-dimensional spaces. This frequently leads to stagnation or premature convergence, necessitating frequent restarts or extensive hyperparameter tuning, which limits their practical utility.

Methodology: The QUASAR Algorithm
To address these limitations, the paper introduces Quasi-Adaptive Search with Asymptotic Reinitialization (QUASAR). Conceptually inspired by the probabilistic search for stability in stellar cores and the behavior of quantum particles, QUASAR extends the DE framework through three highly stochastic mechanisms designed to dynamically balance exploration and exploitation:

Probabilistic Mutation Strategies:
Instead of a fixed mutation strategy, QUASAR employs a "push-and-pull" approach using three distinct strategies selected dynamically based on an "entanglement rate" (default 0.33):
- Spooky-Best (Local/Exploitation): Similar to DE/best/1, this strategy exploits the immediate search space around the current best solution ( $X_{best}$ ) using a mutation factor sampled from a normal distribution $N(0, 0.33^2)$ .
- Spooky-Current & Spooky-Random (Global/Exploration): These strategies drive non-local exploration. They utilize a bimodal distribution for the mutation factor, $N(0.5, 0.25^2) + N(-0.5, 0.25^2)$ , to encourage larger jumps. Notably, the Spooky-Random strategy utilizes the same random vector twice in its formulation.
Rank-Based Crossover:
QUASAR utilizes a dynamic binomial crossover rate ( $CR_i$ ) inversely proportional to the solution's fitness rank. The best solutions (rank 0) are assigned a higher crossover rate (capped at a minimum of 0.33), allowing them to rapidly recombine components from mutant vectors. This ensures that high-quality solutions inherit beneficial traits while maintaining diversity.
Asymptotic Covariance Reinitialization:
A primary innovation of QUASAR is its reinitialization mechanism. In early generations, the worst-performing 33% of the population is reinitialized. The probability of this event decays asymptotically toward a target of 0.33 as generations progress.
- Mechanism: New solutions are sampled from a multivariate normal distribution defined by the mean ( $\mu$ ) and covariance matrix ( $\Sigma$ ) of the top 25% of the population.
- Noise: A noise vector is added to the sampled distribution to further promote exploration.
- Constraint: This mechanism is bypassed if the population size ( $N$ ) is less than the dimensionality ( $D$ ) to prevent singular covariance matrices.

Experimental Design
The efficacy of QUASAR was rigorously evaluated against standard DE (SciPy implementation) and the state-of-the-art L-SHADE (MealPy implementation) using the full CEC2017 benchmark suite of 29 test functions.

Parameters: Experiments utilized default hyperparameters for all algorithms to ensure fair comparison.
Variants: Trials were conducted across varying dimensions ( $D \in [10, 30, 50, 100]$ ) with a fixed population size ( $N=1000$ ), and varying population sizes ( $N \in [100, 5000]$ ) with a fixed dimension ( $D=100$ ).
Metrics: Performance was measured using Friedman rank sums, geometric mean error reduction factors ( $R_{error}$ ), and optimization run time ratios ( $R_{time}$ ). Statistical significance was verified using the Wilcoxon signed-rank test.

Key Results
QUASAR demonstrated superior performance across all metrics compared to both DE and L-SHADE:

Overall Ranking: Using the Friedman test, QUASAR achieved the lowest overall rank sum (367), significantly outperforming L-SHADE (452) and DE (735).
Solution Quality: Geometric mean comparisons revealed that QUASAR improved final solution quality by an average of 3.85× over DE and 2.07× over L-SHADE ( $p \ll 0.001$ ). The most significant gains were observed in higher dimensions (e.g., 100D) and larger population sizes.
Computational Efficiency: QUASAR was 1.40× faster than DE and 5.16× faster than L-SHADE on average. The algorithm scaled linearly with dimensionality, maintaining efficiency even as problem complexity increased.
Robustness: The algorithm performed effectively using its default parameters, requiring no manual hyperparameter tuning.

Significance and Claims
The paper posits that QUASAR establishes a new standard for evolutionary algorithms in complex, high-dimensional, and non-differentiable optimization problems. Its significance lies in:

Effectiveness: It achieves statistically significant improvements in solution quality without requiring extensive parameter tuning.
Efficiency: It offers substantial speedups over current state-of-the-art methods, making it viable for computationally expensive real-world applications.
User-Friendliness: The reliance on default parameters and a streamlined implementation (available as an open-source Python package, hdim_opt) allows for seamless integration into existing workflows.

The authors conclude that QUASAR is a powerful, ready-to-use tool for the next generation of high-dimensional optimization challenges, though they note that future work could explore weighted covariance matrices (similar to CMA-ES) and verification on constrained or multi-objective problems.