Pietro Simone Oliveto

Simplified Drift Analysis for Proving Lower Bounds in Evolutionary Computation

Drift analysis is a powerful tool used to bound the optimization time of evolutionary algorithms (EAs). Various previous works apply a drift theorem going back to Hajek in order to show exponential lower bounds on the optimization time of EAs. However, this drift theorem is tedious to read and to apply since it requires two bounds on the moment-generating (exponential) function of the drift. A recent work identifies a specialization of this drift theorem that is much easier to apply. Nevertheless, it is not as simple and not as general as possible. The present paper picks up Hajek’s line of thought to prove a drift theorem that is very easy to use in evolutionary computation. Only two conditions have to be verified, one of which holds for virtually all EAs with standard mutation. The other condition is a bound on what is really relevant, the drift. Applications show how previous analyses involving the complicated theorem can be redone in a much simpler and clearer way. In some cases even improved results may be achieved. Therefore, the simplified theorem is also a didactical contribution to the runtime analysis of EAs.

Analysis of diversity-preserving mechanisms for global exploration*

Maintaining diversity is important for the performance of evolutionary algorithms. Diversity-preserving mechanisms can enhance global exploration of the search space and enable crossover to find dissimilar individuals for recombination. We focus on the global exploration capabilities of mutation-based algorithms. Using a simple bimodal test function and rigorous runtime analyses, we compare well-known diversity-preserving mechanisms like deterministic crowding, fitness sharing, and others with a plain algorithm without diversification. We show that diversification is necessary for global exploration, but not all mechanisms succeed in finding both optima efficiently. Our theoretical results are accompanied by additional experiments for different population sizes.

Analysis of the

Vertex cover is one of the best known NP-hard combinatorial optimization problems. Experimental work has claimed that evolutionary algorithms (EAs) perform fairly well for the problem and can compete with problem-specific ones. A theoretical analysis that explains these empirical results is presented concerning the random local search algorithm and the (1+1)-EA. Since it is not expected that an algorithm can solve the vertex cover problem in polynomial time, a worst case approximation analysis is carried out for the two considered algorithms and comparisons with the best known problem-specific ones are presented. By studying instance classes of the problem, general results are derived. Although arbitrarily bad approximation ratios of the (1+1)-EA can be proved for a bipartite instance class, the same algorithm can quickly find the minimum cover of the graph when a restart strategy is used. Instance classes where multiple runs cannot considerably improve the performance of the (1+1)-EA are considered and the characteristics of the graphs that make the optimization task hard for the algorithm are investigated and highlighted. An instance class is designed to prove that the (1+1)-EA cannot guarantee better solutions than the state-of-the-art algorithm for vertex cover if worst cases are considered. In particular, a lower bound for the worst case approximation ratio, slightly less than two, is proved. Nevertheless, there are subclasses of the vertex cover problem for which the (1+1)-EA is efficient. It is proved that if the vertex degree is at most two, then the algorithm can solve the problem in polynomial time.

(1+1)

Immune algorithms have been used widely and successfully in many computational intelligence areas including optimization. Given the large number of variants of each operator of this class of algorithms, this paper presents a study of the convergence properties of immune algorithms in general, conducted by examining conditions which are sufficient to prove their convergence to the global optimum of an optimization problem. Furthermore problem independent upper bounds for the number of generations required to guarantee that the solution is found with a defined probability are derived in a similar manner as performed previously, in literature, for genetic algorithms. Again the independence of the function to be optimised leads to an upper bound which is not of practical interest, confirming the general idea that when deriving time bounds for evolutionary algorithms the problem class to be optimised needs to be considered

-EA for Finding Approximate Solutions to Vertex Cover Problems

We investigate theoretically how the fitness landscape influences the optimization process of population-based evolutionary algorithms using fitness-proportional selection. Considering the function OneMax, we show that it cannot be optimized in polynomial time with high probability regardless of the population size. This is proved by a generalization of drift analysis. For populations of at most logarithmic size, the negative result transfers to any function with unique optimum. Based on these insights, we investigate the effect of scaling the objective function in combination with a population that is not too small and show that then such algorithms compute optimal solutions for a wide range of problems in expected polynomial time. Finally, relationships with (1+λ)-EAs and (1,λ)-EAs are described.

On the Convergence of Immune Algorithms

Recently it has been proved that the (1+1)-EA produces poor worst-case approximations for the vertex cover problem. In this paper the result is extended to the (1+lambda)-EA by proving that, given a polynomial time, the algorithm can only find poor covers for an instance class of bipartite graphs. Although the generalisation of the result to the (mu+1)-EA is more difficult, hints are given in this paper to show that this algorithm may get stuck on the local optimum of bipartite graphs as well because of premature convergence. However a simple diversity maintenance mechanism can be introduced into the EA for optimising the bipartite instance class effectively. It is proved that the diversity mechanism combined with one point crossover can change the runtime for some instance classes from exponential to polynomial in the number of nodes of the graph.

Theoretical analysis of fitness-proportional selection: landscapes and efficiency

Maintaining diversity is important for the performance of evolutionary algorithms. Diversity mechanisms can enhance global exploration of the search space and enable crossover to find dissimilar individuals for recombination. We focus on the global exploration capabilities of mutation-based algorithms. Using a simple bimodal test function and rigorous runtime analyses, we compare well-known diversity mechanisms like deterministic crowding, fitness sharing, and others with a plain algorithm without diversification. We show that diversification is necessary for global exploration, but not all mechanisms succeed in finding both optima efficiently.

Analysis of population-based evolutionary algorithms for the vertex cover problem

Parameter setting is an important issue in the design of evolutionary algorithms. Recently, experimental work has pointed out that it is often not useful to work with a fixed mutation rate. Therefore it was proposed that the population be ranked according to fitness and the mutation rate of an individual should depend on its rank. The claim is that this allows the algorithm to explore new regions in the search space as well as progress quickly towards optimal solutions. Complementing the experimental investigations, we examine the proposed approach by presenting rigorous theoretical analyses which point out the differences of rank-based mutation compared to a standard approach using a fixed mutation rate. To this end we theoretically explain the behaviour of rank-based mutation on various fitness landscapes proposed in the experimental work and present new significant classes of functions where the use of rank-based mutation may be both beneficial or detrimental compared to fixed mutation strategies.

Theoretical analysis of diversity mechanisms for global exploration

Experimental results have suggested that evolutionary algorithms may produce higher quality solutions for instances of vertex cover than a very well known approximation algorithm for this NP-complete problem. A theoretical analysis of the expected runtime of the (1+1)-EA on a well studied instance class confirms such a conjecture for the considered class. Furthermore, a class for which the (1+1)-EA takes exponential optimization time is examined. Nevertheless, given polynomial time, the evolutionary algorithm still produces a better solution than the approximation algorithm. Recently, the existence of an instance class has been proved for which the (1+1)-EA produces poor approximate solutions, given polynomial time. Here it is pointed out that, by using multiple runs, the (1+1)-EA finds the optimal cover of each instance of the considered graph class in polynomial time.

Theoretical analysis of rank-based mutation - combining exploration and exploitation

Island models are popular ways of parallelizing evolutionary algorithms as they can decrease the parallel running time at low communication costs and lead to an increased population diversity. This in particular provides a good setting for crossover as this operator relies on a good diversity between parents. We consider the effect of recombining migrants with individuals on the target island. We rigorously prove, for a test function in pseudo-Boolean optimization, exponential performance gaps between island models with strongly connected topologies and a panmictic (mu+1)-EA as long as the migration interval is not too small. We then choose vertex cover as a classical NP-hard problem. By considering instances with a clear building block structure we prove that, also in this more practical setting, island models with a particular topology drastically outperform panmictic populations. Both the theoretical and empirical results show that for strongly connected topologies, such as ring, the performance drops by decreasing the migration interval, while this is not the case for topologies connected weakly such as the single receiver model.