💻 computer science

Dual-Stage Guided Artificial Lemming Algorithm (DALA) For Solving Engineering Application Optimization Problems

This paper proposes the Dual-Stage Guided Artificial Lemming Algorithm (DALA), which enhances the traditional ALA through adaptive probability ranking, a two-stage greedy strategy combining adversarial and topological learning, and JADE mutation to effectively solve high-dimensional engineering optimization problems and hyperparameter tuning for power load forecasting.

Original authors: Jiayi Gao, Yu Zhang

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiayi Gao, Yu Zhang

Original paper licensed under CC BY 4.0 (https://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 absolute best spot in a massive, foggy forest to set up a campsite. You want the view to be perfect, the ground to be flat, and the mosquitoes to be nonexistent. But the forest is so huge and the fog so thick that you can't see the whole picture at once. This is the daily struggle of "optimization," a branch of science where computers try to find the single best solution to a problem among millions of possibilities. In the real world, these problems aren't just about campsites; they are about designing bridges that use the least amount of steel, tuning the brain of a self-driving car, or predicting how much electricity a city will need tomorrow.

To solve these puzzles, scientists often use "metaheuristic algorithms." Think of these as digital explorers that mimic how nature solves problems. Instead of checking every single spot in the forest (which would take forever), these algorithms send out a swarm of virtual agents that wander around, learn from each other, and gradually zero in on the best spot. One popular explorer is the "Artificial Lemming Algorithm" (ALA), which copies the behavior of real lemmings: they migrate long distances, dig burrows, forage for food, and run from predators. But even these digital lemmings have a flaw: sometimes they get stuck in a small, mediocre patch of the forest, thinking it's the best place, and miss the amazing view just over the hill.

This paper introduces a smarter, upgraded version called the Dual-Stage Guided Artificial Lemming Algorithm (DALA). The researchers, Jiayi Gao and Yu Zhang from Northeast Forestry University, realized that the original lemmings were a bit too rigid. They didn't know when to stop wandering and start digging, or when to stick together and when to split up. To fix this, they gave the lemmings a "coach" and a "game plan."

First, they added a Rank-Based Adaptive Probability system. Imagine the lemmings are a sports team. In the old version, everyone got the same instructions regardless of how well they were playing. In DALA, the coach looks at the scoreboard. The top players (the "elite" lemmings who found good spots) are told to focus intensely on refining their immediate area—digging deeper and checking every inch of the ground. The players who are struggling (the "bottom-tier" lemmings) are told to take a big risk and run far away to find a completely new area. This ensures the team doesn't waste time on bad spots while also making sure the best spots are explored thoroughly.

Second, they introduced a Two-Stage Greedy Strategy. This is like a "double-check" system. Before a lemming settles into a new burrow, it doesn't just trust its first instinct. It pauses and asks two questions: "What if I looked at this from the exact opposite angle?" (a technique called Opposition-Based Learning) and "What if I asked my three closest neighbors what they think?" (Topological Neighborhood Learning). If the new idea is better, the lemming takes it. This helps the algorithm escape "local traps"—those mediocre spots that look good but aren't the best.

Finally, they borrowed a trick from a different algorithm called JADE. Sometimes, even the best teams get stuck in a routine. To shake things up, DALA randomly injects a bit of chaos, forcing some lemmings to jump to a totally new, unpredictable location. This keeps the search fresh and prevents the whole group from getting stuck in the same loop.

The researchers tested this new "super-lemming" against eight other popular algorithms using a set of 29 standard, tricky math problems (known as the CEC 2017 and CEC 2020 benchmarks). The results were impressive: DALA found better solutions more often than the others, especially on the hardest, most complex problems. It didn't just win on paper, though; the team also used it to solve five real-world engineering challenges, like designing a pressure vessel (a tank that holds gas under high pressure) and a gear system for a machine. In these tests, DALA found designs that were lighter, cheaper, or more efficient than those found by the other methods.

Perhaps most excitingly, they used DALA to tune the "brain" of a computer model designed to predict electricity usage. By finding the perfect settings for this model, DALA helped it predict power loads with much higher accuracy than before. This suggests that DALA isn't just a theoretical toy; it's a powerful tool that could help engineers build better machines and keep our power grids running smoothly. While the researchers admit the algorithm still has room to grow—especially in how it starts its search and handles certain types of complex, mixed-up problems—their work shows that by giving digital lemmings a better strategy, we can solve some of the world's toughest engineering puzzles more effectively.

Technical Summary: Dual-Stage Guided Artificial Lemming Algorithm (DALA)

Problem Statement
The paper addresses the limitations of the original Artificial Lemming Algorithm (ALA) when applied to high-dimensional, complex, single-objective optimization problems. While ALA, inspired by lemming behaviors (migration, burrowing, foraging, and predator avoidance), offers a metaheuristic framework, it often suffers from confined local search, limited global exploration, and a lack of solution diversity in high-dimensional spaces. These deficiencies lead to premature convergence and an inability to escape local optima, particularly as problem dimensionality increases. The study aims to develop a more robust algorithm capable of balancing exploration and exploitation effectively in engineering and machine learning contexts.

Methodology
The proposed solution is the Dual-Stage Guided Artificial Lemming Algorithm (DALA). This algorithm retains the core behavioral modeling of the original ALA but integrates three distinct enhancement strategies to form a coordinated feedback architecture:

Rank-Based Adaptive Probability Mechanism:
Instead of using fixed parameters (e.g., random thresholds) to determine strategy selection, DALA employs a ranking system based on individual fitness. Individuals are sorted, and a quadratic mapping function ( $P_i = (Rank_i/N)^2$ ) assigns adaptive probabilities. High-performing (elite) individuals receive lower probabilities for exploration, encouraging intensive local exploitation, while low-performing individuals receive higher probabilities for exploration to drive global search. This dynamically allocates computational resources based on population performance.
Two-Stage Greedy Strategy (Opposition-Based Learning & Topological Neighborhood):
To enhance collaborative learning and escape local optima, a "U-shaped" trigger probability mechanism is introduced. This mechanism selectively activates a greedy update process for both elite and bottom-tier individuals.
- Opposition-Based Learning (OBL): Generates opposite solutions to expand the search space.
- Topological Neighborhood Learning: Individuals interact with their best-performing neighbors to guide movement.
- Greedy Selection: If the trigger condition is met, the algorithm compares the current solution, the basic new solution, and the alternative solution (generated via OBL or neighborhood learning), retaining the best. This ensures that computational effort is focused on the most promising updates.
JADE Differential Evolution Hybrid Module:
To prevent population diversity loss and premature convergence, the algorithm integrates the mutation operator from the JADE (Adaptive Differential Evolution with Optional External Archive) algorithm. Specifically, a $(current-to-pbest)$ mutation strategy is used, where the guiding vector is selected from the top 30% of the population. This introduces stochastic perturbations that help individuals deviate from established evolutionary paths, enhancing the ability to explore new regions of the search space.

Key Contributions

Triple Innovation Mechanism: The paper posits that the primary contribution is not merely the addition of individual components, but the implementation of a coordinated feedback architecture. This framework links ranking-based probability allocation with differential evolution and greedy strategies to dynamically balance search tasks.
Adaptive Strategy Selection: The replacement of fixed parameters with a rank-based quadratic probability mechanism allows the algorithm to self-regulate the balance between exploration and exploitation without manual hyperparameter tuning.
Enhanced Local/Global Search: The integration of OBL and topological learning specifically targets the "middle-tier" stagnation and the "bottom-tier" lack of direction, while the JADE module specifically targets the "elite" risk of premature convergence.

Experimental Results
The performance of DALA was evaluated through extensive experiments:

Benchmark Functions: Tests were conducted on the IEEE CEC2017 (29 functions) and CEC2020 (10 functions) suites, covering unimodal, multimodal, hybrid, and composition functions across dimensions of 10, 30, 50, and 100.
- Statistical Performance: DALA achieved the best mean fitness values in 24 out of 29 CEC2017 functions and 7 out of 10 CEC2020 functions.
- Statistical Significance: Friedman and Wilcoxon signed-rank tests confirmed that DALA's performance is statistically significantly superior to the original ALA and seven other competitive algorithms (including PSO, SSA, GWO, and MVO).
Engineering Applications: The algorithm was applied to five constrained engineering design problems:
1. Pressure Vessel Design
2. Three-Bar Truss Design
3. Welded Beam Design
4. Gear Train Design
5. Stepped Pulley Design
  In these tests, DALA consistently found optimal or near-optimal solutions, often outperforming the original ALA and other metaheuristics in terms of cost minimization and constraint satisfaction.
Real-World Application: DALA was used to optimize the hyperparameters of a multi-layer LSTM network for power load forecasting. The optimized model achieved the lowest Root Mean Square Error (RMSE = 2.3405) and Mean Absolute Percentage Error (MAPE = 3.79%) among all tested algorithms, demonstrating DALA's efficacy in high-dimensional, non-linear machine learning tasks.

Significance and Claims
The authors claim that DALA provides a robust solution for complex engineering optimization problems characterized by high dimensionality, non-linearity, and strong constraints. The study emphasizes that the coordinated feedback architecture is the key to its success, allowing the algorithm to adaptively allocate search tasks and maintain diversity.

The paper modestly acknowledges limitations, noting that DALA currently lacks sufficient intervention during the initialization phase, which can lead to slower initial convergence. Furthermore, the sensitivity of the adaptive switching between global and local strategies requires further refinement. The authors state that future work will focus on extending DALA to discrete-continuous mixed variables, large-scale combinatorial optimization, and dynamic environments, as well as applying it to broader industrial scenarios like energy scheduling and structural design. The study concludes that while DALA shows significant promise, its structural universality in these more severe scenarios requires further mathematical verification.

More like this