Introduction
Evolutionary Algorithms (EAs) are robust optimization techniques inspired by biological evolution, designed to navigate complex, high-dimensional search spaces. A fundamental challenge in EAs is the "selection" process—determining which individuals survive to propagate their traits. Standard selection methods, such as Boltzmann selection, often struggle with premature convergence, where the population becomes trapped in local optima. This paper presents a Generalized Evolutionary Algorithm that leverages Tsallis statistics to provide a more flexible and effective selection mechanism.
Methodology
The core innovation of this research is the replacement of the traditional Gibbs-Boltzmann distribution in the selection phase with the Tsallis canonical distribution.
Tsallis-Based Selection
The probability of selecting a solution is governed by the non-extensive index $q$. When $q=1$, the system recovers standard Boltzmann statistics. However, for $q \neq 1$, the distribution exhibits "long-tailed" characteristics, allowing the algorithm to maintain diversity by retaining less fit individuals with higher probability than standard methods. This helps prevent the search from narrowing too quickly.
Implementation Details
The algorithm employs a dynamic $q$ parameter, starting with an initial value $q_0$ and linearly decreasing it toward 1 as generations progress. This transition allows for high exploration in early stages and increased exploitation as the search matures. Additionally, a Cauchy annealing schedule is used to manage selection pressure via the inverse temperature $\beta$.
Experimental Evaluation
The proposed algorithm was tested against three classic benchmark functions: Ackley’s, Rastrigin’s, and Griewangk’s. These functions represent significant optimization challenges due to their vast search spaces (approximately $2^{75}$ states) and numerous local minima.
Results and Discussion
The experimental results demonstrate that the Tsallis-based selection consistently outperforms both standard Boltzmann and proportionate selection across all benchmarks.
As illustrated in the performance plots, the Tsallis-based approach (dotted line) achieves lower fitness values more rapidly. While Boltzmann selection often plateaus early—indicating entrapment in a local minimum—the Tsallis algorithm continues to find improvements. The choice of $q_0$ is critical; for instance, $q_0 = 1.5$ was found optimal for the Ackley function, while $q_0 = 2$ performed best for Rastrigin’s.
Conclusion
By integrating Tsallis statistics into the evolutionary framework, the authors have developed a more powerful optimization tool. The $q$ parameter provides a tunable mechanism to balance exploration and exploitation, effectively mitigating the risk of premature convergence. Future research will focus on automating the tuning of $q_0$ and extending the algorithm to real-valued optimization problems.
Summary updated