Evolution of the Population via PSO: A Memetic-GRASP Breakthrough in Clustering

Executive Summary

TL;DR: This research tackles the dual challenge of feature selection and data clustering by proposing a Memetic-GRASP hybrid. By replacing traditional local search procedures in a Genetic Algorithm with a Particle Swarm Optimization (PSO) phase, the authors enable individuals to "evolve" based on population-wide knowledge. The result is a highly robust clustering framework that consistently outperforms standard metaheuristics like Ant Colony Optimization (ACO) and Tabu Search.

Academic Positioning: This work represents a sophisticated "hybrid metaheuristic" approach. It bridges the gap between global evolutionary search (GA), social intelligence (PSO), and adaptive randomized heuristics (GRASP), establishing a new state-of-the-art for benchmark clustering problems.

Problem & Motivation: The Curse of Clusters

Clustering $N$ objects into $K$ clusters is fundamentally an NP-hard problem. The complexity arises from two fronts:

1. High Dimensionality: Not all features contribute equally to the structure of a cluster. Irrelevant features add noise, making the distance metric (often Euclidean) unreliable.
2. Multimodality: Objective functions, such as the Sum of Squared Errors (SSE), are non-linear. Classic algorithms like K-means frequently get trapped in local optima depending on their initial seeds.

The authors observed that while Genetic Algorithms (GA) are good at global exploration, they are inefficient at fine-tuning solutions. Conversely, local search methods are too "local." They hypothesized that a social evolution mechanism (PSO) could allow GA individuals to improve their traits before selection, mimicking a more realistic biological evolution.

Methodology: The Two-Phase Hybrid

The proposed Memetic-GRASP architecture operates in two distinct but interconnected cycles:

Phase 1: Feature Selection (Memetic-PSO)

The "chromosome" is a binary vector representing the activation of features. Instead of just using crossover and mutation, the algorithm introduces a PSO Evolution Phase:

• Each GA individual is treated as a particle in a swarm.
• Individuals update their velocity and position based on their own best experience ($p_{id}$) and the population's best ($p_{gd}$).
• This "memetic" step ensures that by the time selection occurs, the population has already moved toward high-potential regions of the feature space.

Phase 2: Clustering Solution (GRASP)

To evaluate the fitness of a feature subset, the GRASP algorithm is invoked:

• Construction: It builds an initial clustering solution using a Restricted Candidate List (RCL), balancing greediness with randomness.
• Local Search: It iteratively reassigns samples to clusters to minimize the SSE until stability is reached.

$J(w,z)={\sum }_{i=1}^N{\sum }_{j=1}^K{w}_{ij}||x_i-z_j|{|}^2$

Experiments & Results

The algorithm was tested on 9 benchmark datasets from the UCI repository, ranging from the low-dimensional Iris to the high-dimensional Spambase.

Superior Accuracy

In the Wine dataset, the algorithm achieved a staggering 98.87% accuracy. In the Ionosphere dataset, it reached 86.89%, outperforming traditional PSO-ACO and standard GA-GRASP methods.

Efficiency in Feature Selection

A key highlight is the Spambase result. The Memetic-GRASP found a superior solution using only 32 features, whereas other hybrid models required over 50 features to achieve lower accuracy. This demonstrates the "sparsity" and "relevance" discovery power of the PSO-enhanced GA.

Critical Insight & Conclusion

The Takeaway

The core insight of this paper is that "local improvement" does not have to be local. By using PSO as the local search mechanism within a GA, the algorithm leverages global swarm intelligence to refine local individuals. This creates a "best of both worlds" scenario: the survival-of-the-fittest logic of GAs combined with the collaborative movement of PSOs.

Limitations & Future Work

While highly effective, the computational overhead of running a full PSO and GRASP cycle for every generation of the GA is significant. Future research could explore Adaptive GRASP iterations to reduce cost without sacrificing the 96%+ accuracy rates. This methodology holds great promise for complex real-world tasks like genomic data clustering and financial credit risk assessment.