A modified Covariance Matrix Adaptation Evolution Strategy with adaptive penalty function and restart for constrained optimization
Introduction
In real-world, several problems are modeled as nonlinear optimization problems, where objective functions have to be optimized under some given constraints. In this kind of problem, one wants to find that optimizes which iswhere is the objective (or fitness) function to be optimized, and is an n-dimensional vector of design variables, . This vector can contain mixed variables such as integer, discrete and continuous ones. Each can be bounded by lower and upper limits (box constraints); and are called equality and inequality constraints, being their number m and p, respectively. Both kinds of constraint can be linear or nonlinear.
Over the years, several meta-heuristics have been proposed to solve constrained problems and find global optimum solutions. Among these, Evolutionary Algorithms, such as Genetic Algorithms (GA), see (Coello Coello, 2000, Coello Coello and Mezura-Montes, 2002, Michalewicz, 1996), and Evolution Strategies (ES), see (Mezura-Montes & Coello Coello, 2005), have attracted most of the research interest. Swarm Intelligence has also been investigated as tool for solving constrained optimization problems, see for instance the studies on Particle Swarm Optimization (PSO) (Aguirre et al., 2007, Cagnina et al., 2008, Coelho, 2010, He and Wang, 2007, Pant et al., 2009, Hu et al., 2003), Cuckoo Search (Yang & Deb, 2010), or Artificial Bee Colonies (Brajevic, Tuba, & Subotic, 2011). Another method which is becoming increasingly popular in this field is Differential Evolution (DE): in the last decade, several variants of DE specifically designed for constrained optimization have been proposed, see the papers (Mezura-Montes et al., 2004, Mezura-Montes et al., Mezura-Montes and Palomeque-Ortiz, 2009a, Mezura-Montes and Palomeque-Ortiz, 2009b), or the most recent works (de Melo and Carosio, 2012, Mohamed and Sabry, 2012, Wang and Cai, 2011).
In the excellent comparative study proposed in Mezura-Montes and Lopez-Ramirez (2007), DE outperformed PSO, GA, ES on a broad set of numerical and engineering constrained optimization problems, although it was also observed that PSO tends to reach the feasible region faster, while ES is very effective in improving upon feasible solutions previously found. Motivated by this result, recent studies focused on hybrid techniques, in the attempt of overcoming the specific limitations of each meta-heuristic: for example, Liu, Cai, and Wang (2010) embedded three mutation rules typical of DE (rand/1, current-to-best/1, and rand/2) into a PSO framework to force it jump out of stagnation; in a similar way, He and Wang (2007) hybridized PSO with Simulated Annealing, which was used as local-search to improve upon the best solution of the swarm. Similar ideas were investigated in Sun and Garibaldi, 2010, Ullah et al., 2007, where different kinds of memetic algorithms were introduced, combining a global search scheme with one or more local-search methods. More recently, Pescador Rojas and Coello Coello (2012) presented instead an agent based memetic algorithm including a co-evolution approach.
One of the most popular—and perhaps intuitive—constraint-handling technique known from the literature is the use of a penalty function which worsens the fitness of unfeasible functions obtained during the search (Michalewicz, 1996). According to this strategy, feasible solutions (i.e., solutions belonging to the feasible set, in the following denoted as F) are simply evaluated based on their function value (i.e. the original objective function), while a penalty function is applied to unfeasible solutions to artificially generate a poor fitness. The above scheme can be formulated as:where and obviously indicate the objective and penalty function, respectively. Despite its simplicity, the main drawback of this method lies in the determination of suitable penalty functions, which in turn affects the efficiency of the search. Several specific mechanisms have been designed to solve this issue, see for example (Coelho, 2010, Coello Coello and Mezura-Montes, 2002, He and Wang, 2007, Pant et al., 2009). Other studies focused on self-adapting schemes, see (Coello Coello, 2000, Mezura-Montes et al., Michalewicz, 1996, Noman and Iba, 2008), where the penalty function is adapted during the optimization process. An alternative approach, called filter-based EA (Clevenger, Ferguson, & Hart, 2005), aggregates the constraint violations and tries to solve an equivalent bi-objective problem, minimizing both the original objective function and the aggregated constraint violation function.
It can be easily seen that all these works have in common the underlying idea that population-based meta-heuristics, due to their parallel nature, intrinsically form an efficient tool for solving constrained optimization problems. On the other hand, in order to apply population-based algorithms to this kind of problems, it is imperative to endow them of proper ways to handle constraints, so to avoid unwanted situations such as stagnation and premature convergence, and guide the optimization process towards (and within) the bounds of the feasible region (s) of the search-space.
In this paper, we investigate a well-known meta-heuristic which has proven extremely successful on unconstrained problems, being yet barely studied in constrained optimization, namely the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) (Hansen, Müller, & Koumoutsakos, 2003). The choice of looking at CMA-ES as tool for constrained optimization is mostly due to the fact that this algorithm represents the current state-of-the-art in optimization, although very few studies focus on its application to constrained problems. Therefore we believe that it is of great interest trying to exploit the powerful features of robustness of CMA-ES also in constrained contexts. To overcome the limitations of CMA-ES on constrained problems, (1) we modify the basic CMA-ES framework by applying some simple changes in its stop criteria and sampling mechanism and (2) we introduce an adaptive penalty function within the algorithm, which efficiently handles constraints and dynamically adjusts the penalty during the optimization. In our view, this adaptive penalty scheme represents the main contribution of this paper. Interestingly, this penalty scheme could be applied also to meta-heuristics other than CMA-ES. However, since our purpose is to extend the range of applications of CMA-ES to constrained optimization, here we prove the applicability of this penalty function only on CMA-ES, and we assess the advantages due to its use. Extensive numerical experiments and comparisons with state-of-the-art methods complete this paper, where we evaluate the performance of the proposed modified CMA-ES with adaptive penalty function on a broad set of benchmark problems and engineering applications take from specialized literature.
The paper is organized as follows: in Section 2, we briefly introduce the basic CMA-ES algorithm. Section 3 describes the proposed algorithm based on modified CMA-ES, with particular focus on the adaptive penalty function. Section 4 includes the description of the test problems, the details of the experiments, the results obtained, and the related discussion. Finally, in Section 5 concludes this work.
Section snippets
Covariance Matrix Adaptation Evolution Strategy
The Covariance Matrix Adaptation Evolution Strategy, introduced in Hansen et al. (2003), is a variant of classic Evolution Strategies (ES) (Rechenberg, 1971, Schwefel, 1965) which makes use of a distribution model of the population in order to learn the variable linkages and speed up the evolutionary process. CMA-ES consists of the following. At the beginning of the optimization, a mean vector is randomly initialized inside the problem bounds , where n is the number of
The proposed modified CMA-ES scheme
As an alternative to the mechanisms mentioned above, we propose here a modified version of CMA-ES obtained perturbing the original CMA-ES framework by introducing simple modifications into its original implementation (see Hansen et al., 2003). Such modifications are described in Section 3.1. Additionally, we introduce into the algorithmic structure an original adaptive constraint handling technique, presented in Section 3.2.
Numerical results
To evaluate and exemplify the performance of the proposed modified CMA-ES, we consider here the 24 constrained optimization problems defined in the CEC 2006 test suite (Liang et al., 2006) (named g01 …g24) and five engineering design problems taken from the literature (see A). In the following, we first describe our algorithm setup (i.e., the parameter setting, see Section 4.1). Then, we report a quantitative (statistical) analysis of the numerical results obtained with the proposed modified
Conclusions
In this paper we proposed a modified Covariance Matrix Adaptation Evolution Strategy specifically tailored for global constrained optimization. Compared to the standard CMA-ES, the proposed method presents simple but significant changes in the solution sampling mechanism and in the stop criteria. In addition to that, it integrates an adaptive penalty function which dynamically changes the penalty depending on the amplitude of the constraint violations and the fitness of the best feasible
Acknowledgments
INCAS3 is co-funded by the Province of Drenthe, the Municipality of Assen, the European Fund for Regional Development and the Ministry of Economic Affairs, Peaks in the Delta. Vinícius Veloso de Melo would like to thank the financial support given by CNPq (Projeto Universal no. 486950/2013-1).
References (82)
Use of a self-adaptive penalty approach for engineering optimization problems
Computers in Industry
(2000)- et al.
Constraint-handling in genetic algorithms through the use of dominance-based tournament selection
Advanced Engineering Informatics
(2002) Gaussian quantum-behaved particle swarm optimization approaches for constrained engineering design problems
Expert Systems with Applications
(2010)- et al.
An effective co-evolutionary particle swarm optimization for constrained engineering design problems
Engineering Applications of Artificial Intelligence
(2007) - et al.
Constrained optimization based on modified differential evolution algorithm
Information Science
(2012) - et al.
Hybridizing particle swarm optimization with differential evolution for constrained numerical and engineering optimization
Applied Soft Computing
(2010) - et al.
A hybrid particle swarm optimization with a feasibility-based rule for constrained optimization
Applied Mathematics and Computation
(2007) - et al.
An adaptive penalty based covariance matrix adaptation–evolution strategy
Computers & Operations Research
(2013) - et al.
Investigating multi-view differential evolution for solving constrained engineering design problems
Expert Systems with Applications
(2013) - et al.
Hybrid Nelder–Mead simplex search and particle swarm optimization for constrained engineering design problems
Expert Systems with Applications
(2009)
Teaching-learning-based optimization: A novel method for constrained mechanical design optimization problems
Computer-Aided Design
Mine blast algorithm: A new population based algorithm for solving constrained engineering optimization problems
Applied Soft Computing
Differential evolution with dynamic stochastic selection for constrained optimization
Information Sciences
The particle swarm optimization algorithm: convergence analysis and parameter selection
Information Processing Letters
An effective co-evolutionary differential evolution for constrained optimization
Applied Mathematics and Computation
Ockham’s Razor in memetic computing: three stage optimal memetic exploration
Information Sciences
Parallel memetic structures
Information Sciences
An improved group search optimizer for mechanical design optimization problems
Progress in Natural Science
A simple multimembered evolution strategy to solve constrained optimization problems
IEEE Transactions on Evolutionary Computation
Solving engineering optimization problems with the simple constrained particle swarm optimizer
Informatica
Low discrepancy initialized particle swarm optimization for solving constrained optimization problems
Fundamental Information
Engineering optimisation by cuckoo search
International Journal of Mathematical Modelling and Numerical Optimisation
Performance of the improved artificial bee colony algorithm on standard engineering constrained problems
International Journal of Mathematics And Computers In Simulation
Simple feasibility rules and differential evolution for constrained optimization
Self-adaptive and deterministic parameter control in differential evolution for constrained optimization
Constrained evolutionary optimization by means of ()-differential evolution and improved adaptive trade-off model
Evolutionary Computation
Evaluating differential evolution with penalty function to solve constrained engineering problems
Expert Systems with Applications
Accelerating differential evolution using an adaptive local search
IEEE Transactions on Evolutionary Computation
A filter-based evolutionary algorithm for constrained optimization
Evolutionary Computation
Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (CMA-ES)
Evolutionary Computation
Cited by (56)
CMA-ES with exponential based multiplicative covariance matrix adaptation for global optimization
2023, Swarm and Evolutionary ComputationMultiple dynamic penalties based on decomposition for constrained optimization
2022, Expert Systems with ApplicationsAn improved Jaya optimization algorithm with Lévy flight
2021, Expert Systems with ApplicationsCitation Excerpt :In the past few years, TLBO has inspired many other algorithms based on similar principles, among which the so-called Jaya algorithm (Rao, 2016), which also does not need any other parameter besides the population size and the number of generations, has lately shown promising results on various kinds of optimization problems. To overcome this issue, one obvious solution is to restart the algorithm according to some criteria (Caraffini et al., 2013; de Melo & Iacca, 2014). However, one could also think of “embedding” the restart mechanism directly into the way random numbers are sampled during the algorithm execution: in order to do that, recent works have proposed the use of different distributions with different properties (e.g., kurtosis and skewness) and therefore different effects on the algorithmic functioning.
Search and rescue optimization algorithm: A new optimization method for solving constrained engineering optimization problems
2020, Expert Systems with ApplicationsLow-Memory Matrix Adaptation Evolution Strategies Exploiting Gradient Information and Lévy Flight
2024, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)