Elizabeth Montero

A beginner's guide to tuning methods

Metaheuristic methods have been demonstrated to be efficient tools to solve hard optimization problems. Most metaheuristics define a set of parameters that must be tuned. A good setup of that parameter values can lead to take advantage of the metaheuristic capabilities to solve the problem at hand. Tuning strategies are step by step methods based on multiple runs of the metaheuristic algorithm. In this study we compare four automated tuning methods: F-Race, Revac, ParamILS and SPO. We evaluate the performance of each method using a standard genetic algorithm for continuous function optimization. We discuss about the requirements of each method, the resources used and quality of solutions found in different scenarios. Finally we establish some guidelines that can help to choose the more appropriate tuning procedure.

A new algorithm for reducing metaheuristic design effort

The process of designing a metaheuristic is a difficult and time consuming task as it usually requires tuning to find the best associated parameter values. In this paper, we propose a simple tuning tool called EVOCA which allows unexperimented metaheuristic designers to obtain good quality results without have a strong knowledge in tuning methods. The simplicity here means that the designer does not have to care about the initial settings of the tuner. We apply EVOCA to a genetic algorithm that solves NK landscape instances of various categories. We show that EVOCA is able to tune both categorical and numerical parameters allowing the designer to discard ineffective components for the algorithm.

A graph-based immune-inspired constraint satisfaction search

We propose an artificial immune algorithm to solve constraint satisfaction problems (CSPs). Recently, bio-inspired algorithms have been proposed to solve CSPs. They have shown to be efficient in solving hard problem instances. Given that recent publications indicate that immune-inspired algorithms offer advantages to solve complex problems, our main goal is to propose an efficient immune algorithm which can solve CSPs. We have calibrated our algorithm using relevance estimation and value calibration (REVAC), which is a new technique recently introduced to find the parameter values for evolutionary algorithms. The tests were carried out using randomly generated binary constraint satisfaction problems and instances of the three-colouring problem with different constraint networks. The results suggest that the technique may be successfully applied to solve CSPs.

An evaluation of off-line calibration techniques for evolutionary algorithms

Most metaheuristics define a set of parameters that must be tuned. A good setup of those parameter values can lead to take advantage of all the metaheuristic capabilities to solve the problem at hand. Tuning techniques are step by step methods based on multiple runs of the algorithm. In this study we compare three automated tuning methods: F-Race, Revac and ParamILS. We evaluate the performance of each method using a genetic algorithm for combinatorial optimization. The differences and advantages of each technique are discussed. Finally we establish some guidelines that might help to choose a tuning process to use.

Self-calibrating strategies for evolutionary approaches that solve constrained combinatorial problems

In this paper, we evaluate parameter control strategies for evolutionary approaches to solve constrained combinatorial problems. For testing, we have used two well known evolutionary algorithms that solve the Constraint Satisfaction Problems GSA and SAW. We contrast our results with REVAC, a recently proposed technique for parameter tuning.

Calibrating strategies for evolutionary algorithms

The control of parameters during the execution of evolutionary algorithms is an open research area. In this paper, we propose new parameter control strategies for evolutionary approaches, based on reinforcement learning ideas. Our approach provides efficient and low cost adaptive techniques for parameter control. Moreover, it is a general method, thus it could be applied to any evolutionary approach having more than one operator. We contrast our results with tuning techniques and HaEa a random parameter control.

New requirements for off-line parameter calibration algorithms

The process of designing an evolutionary algorithm requires the definition of an adequate representation, a set of components as operators, parameters and parameters values. In practice, some operators can not be helping the evolutionary algorithm to perform his work, thus we require to be able to detect these situations. In this paper we are interested on analyzing the capabilities of off-line calibration techniques, originally designed to tune parameter values, to recognize situations more related to the design process of an evolutionary algorithm. We experimentally analyze the results of the off-line calibration techniques on some specific situations. For that purpose we use some specially designed operators for an evolutionary algorithm which solves the traveling salesman problem.

Improving MMAS using parameter control

Tunning parameters values in metaheuristics is a time consuming task. Techniques to control parameters during the execution have been successfully applied into evolutionary algorithms. The key idea is that the algorithm themselves computes its parameters values according to its current state of the search. In this paper, we propose a strategy to include parameters control on ants based algorithms. We have tested our approach to solve hard instances of the travel salesman problem using MMAS. The tests shown that in some cases, it is possible to obtain better results than the reported ones for the same algorithm, by including a parameter control strategy.

An Overview of Weighted and Unconstrained Scalarizing Functions

Scalarizing functions play a crucial role in multi-objective evolutionary algorithms MOEAs based on decomposition and the R2 indicator, since they guide the population towards nearly optimal solutions, assigning a fitness value to an individual according to a predefined target direction in objective space. This paper presents a general review of weighted scalarizing functions without constraints, which have been proposed not only within evolutionary multi-objective optimization but also in the mathematical programming literature. We also investigate their scalability upi¾źto 10 objectives, using the test problems of Lame Superspheres on the MOEA/D and MOMBI-II frameworks. For this purpose, the best suited scalarizing functions and their model parameters are determined through the evolutionary calibrator EVOCA. Our experimental results reveal that some of these scalarizing functions are quite robust and suitable for handling many-objective optimization problems.

Ants can Learn from the Opposite

In this work we present different learning strategies focused on detecting candidate solutions that are not interesting to be explored by a metaheuristic, in terms of evaluation function. We include a first step before the metaheuristic. The information obtained from this step is given to the metaheuristic, for visiting candidate solutions that are more promising in terms of their quality. The goal of using these strategies is to learn about candidate solutions that can be discarded from the search space, and thus to improve the search of the metaheuristic. We present two new strategies that differ on how the solutions can be constructed in an opposite way. Our approach is evaluated using Ant Solver, a well-known ant based algorithm for solving Constraint Satisfaction Problems. We show promising results that make our solution as good approach to apply in other metaheuristics.