Runtime Analysis of a Multi-Valued Compact Genetic Algorithm on Generalized OneMax

Sumit Adak,Carsten Witt

2024-04-17

Abstract:A class of metaheuristic techniques called estimation-of-distribution algorithms (EDAs) are employed in optimization as more sophisticated substitutes for traditional strategies like evolutionary algorithms. EDAs generally drive the search for the optimum by creating explicit probabilistic models of potential candidate solutions through repeated sampling and selection from the underlying search space.

Neural and Evolutionary Computing

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to analyze the running time of the multi - valued compact genetic algorithm (r - cGA) on the generalized OneMax problem. Specifically, the researchers hope to understand how r - cGA can effectively optimize problems containing multi - valued decision variables and provide theoretical running - time bounds.

### Research Background

Traditional Estimation - of - Distribution Algorithms (EDAs) are mainly used for pseudo - Boolean optimization problems, that is, problems dealing with binary decision variables. However, in practical applications, many problems may involve decision variables with multiple values (i.e., multi - valued decision variables). For this reason, Jedidia et al. first proposed EDAs suitable for multi - valued decision variables at GECCO 2023 and analyzed the running time of the multi - valued UMDA (univariate marginal distribution algorithm) on the r - valued LeadingOnes function by constructing a framework.

### Research Objectives

This paper further explores the performance of r - cGA on the generalized OneMax function on this basis. The generalized OneMax function has two variants: r - OneMax and G - OneMax. Among them:

- The **r - OneMax** function is defined as:
\[
r\text{-OneMax}(x)=\sum_{i = 1}^{n}1\{x_{i}=r - 1\}
\]
That is, it calculates the number of positions in the string equal to \(r - 1\).

- The **G - OneMax** function is defined as:
\[
G\text{-OneMax}(x)=\sum_{i = 1}^{n}x_{i}
\]
That is, it calculates the sum of the values of all positions in the string.

### Main Contributions

1. **Running - time Analysis**: The author proves that the running - time bound of r - cGA on the r - OneMax problem is \(O(r^{2}n\log^{2}r\log^{3}n)\) and holds with high probability.

2. **Genetic Drift Analysis**: By analyzing genetic drift, ensure that the algorithm parameter settings are reasonable and avoid performance losses caused by random fluctuations.

3. **Comparison and Conjecture**: Compare the obtained results with those of the binary cGA and propose a conjecture about the expected running time of another multi - valued OneMax variant (G - OneMax).

### Method Overview

To achieve the above objectives, the researchers adopted the following methods:

- **Probability Model Update**: r - cGA represents the probability of each position taking a specific value by marginal probability (frequency) and updates these frequencies according to the fitness of the sampled individuals.

- **Random Walk and Biased Step**: Analyzed two different types of steps in r - cGA - random walk step (rw - step) and biased step (b - step), and gave their probability change laws respectively.

- **Genetic Drift Control**: Use the martingale concentration inequality to control the influence of genetic drift and ensure that the frequency does not deviate too far from the optimal solution due to random fluctuations.

### Conclusion

Through strict theoretical analysis, the researchers prove the efficiency of r - cGA on the multi - valued OneMax problem and provide a new perspective and tool for further research on the optimization problems of multi - valued decision variables.