Fixed-Parameter Tractability of the (1+1) Evolutionary Algorithm on Random Planted Vertex Covers

Jack Kearney,Frank Neumann,Andrew M. Sutton

DOI: https://doi.org/10.1145/3594805.3607134

2024-09-16

Abstract:We present the first parameterized analysis of a standard (1+1) Evolutionary Algorithm on a distribution of vertex cover problems. We show that if the planted cover is at most logarithmic, restarting the (1+1) EA every $O(n \log n)$ steps will find a cover at least as small as the planted cover in polynomial time for sufficiently dense random graphs $p > 0.71$. For superlogarithmic planted covers, we prove that the (1+1) EA finds a solution in fixed-parameter tractable time in expectation. We complement these theoretical investigations with a number of computational experiments that highlight the interplay between planted cover size, graph density and runtime.

Neural and Evolutionary Computing

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper explores the parameterized complexity of the standard (1 + 1) Evolutionary Algorithm (EA) on the randomly planted vertex - cover problem. Specifically, the authors study the following problems:

1. **Fixed - parameter tractability**: When the size \(k\) of the planted vertex - cover in the graph is small (logarithmic level or smaller), can the (1 + 1) EA find a solution that is at least as small as the planted vertex - cover within polynomial time?
2. **Super - logarithmic - level vertex - cover**: When the size \(k\) of the planted vertex - cover is large (super - logarithmic level), can the (1 + 1) EA find a solution within the expected fixed - parameter tractable time?

#### Main contributions

- **Theoretical analysis**:
- For sufficiently dense random graphs (edge probability \(p>0.71\)), if the size \(k\) of the planted vertex - cover is \(k\leq\ln n\), by restarting the (1 + 1) EA every \(O(n\log n)\) steps, a solution that is at least as small as the planted vertex - cover can be found within polynomial time.
- For larger \(k\), it is proved that the expected running time of the (1 + 1) EA is \(O(f(k, p)n\log n)\), where \(f\) is a function of \(k\) and \(p\), but does not depend on \(n\).

- **Computational experiments**:
- A series of computational experiment results are provided to explore the relationships among the size \(k\) of the planted vertex - cover, the graph density \(p\), and the running time.
- The experimental results show that within certain parameter ranges, there are significant differences in the performance of the (1 + 1) EA, suggesting new directions for future theoretical research.

#### Formula summary

For smaller \(k\) (i.e., \(k\leq\ln n\)), if the edge probability \(p\) satisfies:
\[p > \sqrt{\frac{1-\ln\delta}{2}}\]
then the probability that the (1 + 1) EA finds a \(k\)-cover within \(O(n^{c + 1}\log n)\) time is:
\[\Pr(T\leq 2en\ln n+\lfloor en(1 - \delta)\rfloor)=\Omega(n^{-(e(1-\delta)\ln(2e)+\ln 2)})\]

For larger \(k\) (i.e., \(k > \ln n\)), the expected running time of the (1 + 1) EA is:
\[O\left(k^4k\left(1+\frac{1}{p}\right)n\log n\right)\]

These results indicate that the (1 + 1) EA has fixed - parameter tractability when dealing with the randomly planted vertex - cover problem, and in particular, it performs well when the graph density is high and the size of the planted vertex - cover is moderate.