An introduction to synchronous self-learning Pareto strategy

Ahmad Mozaffari, Alireza FathiarXiv 2013 · arXiv

Introduction

In modern engineering design, objectives are rarely singular. Whether maximizing the structural integrity of a bridge while minimizing material costs, or optimizing an engine for both power and fuel efficiency, engineers face multi-objective optimization (MOO) problems. Because these objectives often conflict, there is no single "perfect" solution. Instead, researchers seek the Pareto front—a set of solutions where no objective can be improved without degrading another.

The effectiveness of an MOO algorithm is measured by two primary criteria: convergence (how close solutions are to the true theoretical limit) and diversity (how well-distributed solutions are across that limit). While algorithms like NSGA-II have long been the industry standard, they can struggle with robustness in complex landscapes. This paper introduces the Synchronous Self-Learning Pareto Strategy Algorithm (SSLPSA), a hybrid approach designed to overcome these limitations by fusing evolutionary computation, swarm intelligence, and neural networks.

The SSLPSA Framework

SSLPSA is a multi-layered system that operates synchronously to balance global exploration with local refinement. It consists of three core components:

1. Evolutionary and Swarm Components

The algorithm utilizes a Tournament Based Genetic Algorithm (TBGA) for broad exploration of the search space through selection and crossover. To complement this, a Quasi Artificial Bee Colony (QABC) algorithm is employed. In a departure from standard ABC implementations, the authors removed the "scout bee" phase. This modification transforms the swarm component into a specialized tool for intensification, focusing on polishing solutions in promising local areas rather than random exploration.

2. Self-Learning via SOM

A unique feature of SSLPSA is the integration of Self-Organizing Maps (SOM). This unsupervised neural network "learns" the topology of the Pareto front in real-time. To prevent the network from stagnating, the authors implemented a "conscience mechanism."

The conscience mechanism in the SOM prevents units from becoming 'dead' and ensures continuous learning of the search space.

As shown in Figure 2, the conscience mechanism monitors neuron activity. If a neuron (or "unit") becomes inactive—a "dead unit"—the mechanism adjusts its bias to revive it, ensuring that the entire network contributes to identifying features of the Pareto front and maintaining solution diversity.

3. Synchronous Shuffling

The "Synchronous" aspect of SSLPSA refers to the Sharing Factor ($\xi$), a mechanism that shuffles data between the Swarm and Genetic components. This allows the system to remain robust; if one optimization path becomes trapped in a local optimum, the exchange of high-quality solutions from the other path can provide the necessary "push" to escape.

Experimental Validation and Results

The performance of SSLPSA was evaluated against seven mathematical benchmarks (including the ZDT series, SCH, and FON) and compared with established algorithms like NSGA-II and SPEA 2.

Convergence and Diversity

SSLPSA demonstrated superior convergence in 5 out of 7 benchmarks. Specifically, it achieved better Inverted Generational Distance (IGD) and SPREAD values, indicating that its solutions were both closer to the true front and more evenly distributed. For the ZDT1 problem, the algorithm's focus on intensification allowed it to identify nearly 3,500 non-dominated solutions in a single run.

The non-dominated sorting procedure categorizes solutions into successive fronts to guide the search toward the optimal Pareto set.

The algorithm relies on a non-dominated sorting procedure (Figure 4) to rank solutions. By categorizing the population into different "fronts," the algorithm prioritizes the most promising trade-offs while maintaining a pool of diverse candidates for future generations.

Discussion and Future Work

While SSLPSA offers significant improvements in robustness and solution volume, it is computationally more intensive than simpler models like NSGA-II due to its multi-layered architecture. The authors noted that while it excelled in most benchmarks, it faced challenges with specific landscapes (like FON and ZDT3) where local convergence occasionally hindered global discovery.

Future research will focus on transitioning to an asynchronous parallel model to reduce CPU time and exploring the integration of non-cooperative game theory to further refine agent interactions within the hybrid framework.

Conclusion

SSLPSA represents a significant step forward in hybrid optimization. By synchronizing evolutionary strategies with a self-learning neural network, it provides a robust framework for tackling the complex, conflicting objectives inherent in modern engineering.

Summary updated

BibTeX
@misc{mozaffari2013introduction,
  author = {Ahmad Mozaffari and Alireza Fathi},
  title = {{An introduction to synchronous self-learning Pareto strategy}},
  year = {2013},
  eprint = {1312.4132},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1312.4132},
}
Figure