Runtime analysis of evolutionary algorithms via symmetry arguments
Introduction
The theory of evolutionary algorithms (EAs) has produced a decent number of mathematically proven runtime analyses. They explain the working principles of EAs, advise how to use these algorithms and how to choose their parameters, and have even led to the invention of new algorithms. We refer to [1], [3], [9], [10] for introductions to this area.
Due to the complexity of the probability space describing a run of many EAs, the majority of the runtime analyses regard very simple algorithms. In particular, there are only relatively few works discussing algorithms that employ crossover, that is, the generation of offspring from two parents. Among these, again very few present lower bounds on runtimes; we are aware of such results only in [7], [12], [14].
In the most recent of these works, Sutton and Witt [14, Section 3] consider a simple crossover-based algorithm called StSt GA0 (made precise in Section 2 below). This steady-state genetic algorithm uses a two-parent two-offspring uniform crossover as only variation operator. The two offspring always replace their parents. There is no fitness-based selection and no mutation in this simple process. Clearly, an algorithm of this kind is not expected to be very useful in practice. The reason to study such algorithms is rather that they allow to analyze in isolation how crossover works (more reasons to study this particular algorithm are described in [14]).
Without fitness-based selection, and thus without regarding the problem to be optimized, one would expect that this algorithm takes an exponential time to find any particular search point of the search space . Surprisingly, this is not so obvious, at least not when working with a particular initialization of the population. Sutton and Witt [14, Theorem 10] initialize the algorithm with copies of the string and copies of the string . They argue that this is a population with extremely high diversity, which could thus be beneficial for a crossover-based algorithm. Sutton and Witt show that their algorithm with this initialization and with population size , a constant, takes an expected number of iterations to generate the target string . Apparently, this lower bound is subexponential for all population sizes. It becomes weaker with increasing population size and is trivial for .
By exploiting symmetries in the stochastic process, we improve the lower bound to for all values of μ.
Theorem 1 Let with μ and n even. Consider a run of the StSt GA0 initialized with copies of and copies of . Then the probability that the target string is generated in the first t iterations is at most . In particular, the expected time to generate is at least .
Our proof is based on a simple group action or symmetry argument. We observe that the automorphisms of the hypercube (viewed as graph) commute with the operations of the StSt GA0. Consequently, if an automorphism σ stabilizes the initial individuals z and (that is, and ), then for any at all times t the probability that the algorithm generates x equals the probability that it generates .
From this symmetry, we conclude that if B is the set of all x such that there is an automorphism of the hypercube that stabilizes the initial individuals and such that , then at all times the probability that is generated, is at most . We compute that B has exactly elements. Hence each search point generated by the StSt GA0 is equal to only with probability . A simple union bound over the 2t search points generated up to iteration t gives the result.
Section snippets
Precise problem statement
The algorithm regarded in [14, Section 3], called StSt GA0, is a selection-free variant of a steady state genetic algorithm proposed earlier in [15]. It works with the search space of bit strings of length n, which is a standard representation used in evolutionary computation. The algorithm uses a population of size . Each iteration consists of (i) choosing two different individuals randomly from the population, (ii) applying a two-offspring uniform crossover, and (iii)
Proof of the main result
We now prove our main result following the outline given towards the end of Section 1. We do not assume any prior knowledge on groups and their action on sets.
We view the hypercube as a graph in the canonical way, that is, two bit strings are neighbors if and only if they differ in exactly one position, that is, if . A permutation σ of is called graph automorphism if it preserves the neighbor relation, that is, if x and y are neighbors if and only if and
Conclusion and open problems
We proposed an alternative approach to the problem how long the StSt GA0 with a particular initialization takes to generate a particular search point [14, Section 3.1]. Our lower bound of order , valid for all population sizes μ, is significantly stronger than the previous result, which is at most and decreases with increasing population size until it is trivial for . Our main argument based on group actions is elementary and natural, which gives us the hope that
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work was supported by a public grant as part of the Investissement d'avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH.
References (15)
- et al.
Improved time complexity analysis of the simple genetic algorithm
Theor. Comput. Sci.
(2015) - et al.
The impact of random initialization on the runtime of randomized search heuristics
Algorithmica
(2016) Exponential upper bounds for the runtime of randomized search heuristics
Probabilistic tools for the analysis of randomized optimization heuristics
- et al.
Money for nothing: speeding up evolutionary algorithms through better initialization
Cited by (3)
Finding Symmetry Groups of Some Quadratic Programming Problems
2022, Numerical MathematicsRuntime analysis via symmetry arguments: (hot-off-the-press track at GECCO 2021)
2021, GECCO 2021 Companion - Proceedings of the 2021 Genetic and Evolutionary Computation Conference Companion