CC BY
22Решетневскуе чтения. 2014
УДК 519.8
САМОКОНФИГУРИРУЕМЫЙ АЛГОРИТМ УПРАВЛЕНИЯ СТРАТЕГИЯМИ ЭВОЛЮЦИОННОГО ПОИСКА ДЛЯ НЕСТАЦИОНАРНЫХ ЗАДАЧ ОПТИМИЗАЦИИ
Е. А. Сопов
Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева Российская Федерация, 660014, г. Красноярск, просп. им. газ. «Красноярский рабочий», 31
Е-mail: es_gt@mail.ru
Задачи нестационарной оптимизации являются сложными для многих поисковых алгоритмов. В прикладных задачах оптимизации нет возможности предварительно определить соответствующий алгоритм для конкретной задачи. Представлен оригинальный самоконфигурируемый алгоритм, позволяющий адаптировать комбинацию поисковых стратегий в ходе решения задачи оптимизации.
Ключевые слова: нестационарная оптимизация, динамическая оптимизация, генетические алгоритмы, самоконфигурация.
SELF-CONFIGURING EVOLUTIONARY SEARCH STRATEGY DESIGN FOR NON-STATIONARY OPTIMIZATION PROBLEM
E. A. Sopov
Siberian State Aerospace University named after academician M. F. Reshetnev 31, Krasnoyarsky Rabochy Av., Krasnoyarsk, 660014, Russian Federation Е-mail: es_gt@mail.ru
The non-stationary optimization problem is a great challenge for many search algorithms. In recent years, a variety of different search strategies based on genetic algorithms is developed. Unfortunately, in real-world optimization problems there is no possibility to pre-identify the correct algorithm for the given problem. An original self-configuring approach, which adaptively form search strategies combination during optimization process, is presented.
Keywords: non-stationary optimization, DOP, genetic algorithms, self-configuring.
Many real-world optimization problems are subject to changing conditions over time, so being able to optimize in a dynamic environment is important. Changes may affect the object functions and/or constraints due to the arrival of new tasks, the change of some conditions and external factors and the variance of available resources.
Hence, the optimal solution of the problem may change over time. In the literature of optimization in dynamic environments, researchers usually define optimization problems that change over time as non-stationary, dynamic optimization problems (DOP) or time-dependent problems.
Optimization or adaptation in non-stationary environments is a fairly new research area from the perspective of a computer scientist or engineer. However, from a biologist's viewpoint, adaptation in a changing natural environment is a common theme. That's why there is a good idea to use some population-based and evolutionary approaches to solve DOP. Nowadays there exist a variety of particular DOP solution including ant colony optimization, cooperative strategies, cultural algorithms, evolutionary algorithms, self-organizing scouts, EDA, immune-based algorithms, PSO and other [1; 2]. However, one can say that genetic algorithms are the most common and well-investigated technique.
Many existing classifications of DOP types describe the following problems: altering (or cycling) problems, problems with changing morphology, problems with
drifting landscapes, abrupt and discontinuous problems and other. Thus, there are many search strategies, which are classified using the following distinction:
- restarting techniques where the optimization is started again from scratch;
- local variation to adapt to changes in the environment;
- memorizing techniques where previous solutions (or parts of previous solutions) are reintroduced into the population;
- diversity preserving techniques to avoid the loss of adaptability;
- adaptive and self-adaptive techniques;
- algorithms with overlapping generations;
- learning of the dynamics rules;
- non-local encoding [3].
The implementation of DOP strategies in real-world applications associated with two problems as one has no a priori information about DOP class:
- which of the algorithms to choose;
- what the parameters of genetic algorithm (selection type, crossover/mutation type and rate and other) to set.
The in-run tuning of the certain genetic algorithms is enough studied problem and can be solved using many technique like co-evolution, distribution estimation, self-tuning and self-adaptation and so on. At the same time, there is no efficient way to predefine an appropriate strategy.
Математические методы моделирования, управления и анализа данных
In this work, a new self-configuring genetic algorithm for DOP, which combines many DOP strategies and adaptively control their joint work, is proposed.
In previous works, the SelfCOMOGA algorithm for multi-objective (MO) optimization problem was introduced [4]. This algorithm combines some of MO techniques using the hybrid of competitive and cooperative co-evolution schemes. The DOP algorithm idea of different search strategies cooperation is taken from SelfCOMOGA.
The self-configuring DOP algorithm combine all mentioned above search strategies. An additional strategy can be included in the algorithm with no changes in the algorithm structure. The population size defines the amount of candidate solutions, which are summary evaluated in a moment of time.
As a priory information about types, frequencies and strength of environment change is absent, the initial population is distributed equally between strategies. When the environment change occurs and some number of algorithm iterations (or generations) is done, the performance of the search strategies is estimated. The percentage of population that is handled by "efficient" at the current time algorithm should be increased. These individuals are taken from less efficient ones. This step can be viewed as competitive co-evolution. It's important to save some minimum threshold percentage to give every algorithm a chance to prove his search abilities with a next environment changes.
When a new distribution of population is formed, some random migrations of the best individuals are performed. This step can be viewed as cooperative co-evolution.
The fine tuning of the individual search strategies is performed using the self-configuring genetic algorithm idea developed in [5]. In that work the variation of genetic operations implementation is defined by probabilities distribution, which continuously re-estimated according to success of the certain operation result. Therefore, there is no need to control the individual algorithm parameters in co-evolution scheme.
The algorithm efficiency was investigated solving a common DOP test problem (the good survey can be found in [1]). It demonstrates the better performance on average than the average performance over individual strategies. The detailed results will be performed in the conference presentation.
The main advantage of the proposed approach is than the DOP problem is solved with no additional information
УДК 519.6
about problem. The search strategies are adaptively controlled by algorithm and form a kind of an optimal interaction structure at each moment of optimization process. So the algorithm can be named as self-configuring.
Библиографические ссылки
1. Cruz C., González J. R., Pelta D. Optimization in dynamic environments: a survey on problems, methods and measures // Soft Computing. 2012. 15 (7).
2. Nguyena T. T., Yang S., Branke J. Evolutionary dynamic optimization: A survey of the state of the art // Swarm and Evolutionary Computation. 2012. 6.
3. Weicker K. Evolutionary algorithms and dynamic optimization problems // Der Andere Verlag, 2003.
4. Иванов И. А., Сопов Е. А. Исследование эффективности самоконфигурируемого коэволюционного алгоритма решения сложных задач многокритериальной оптимизации // Системы управления и информационные технологии. 2013. № 1.1 (51).
5. Semenkin E., Semenkina M. Self-configuring genetic algorithm with modified uniform crossover operator // ADVANCES IN SWARM INTELLIGENCE (ICSI'2012). LNCS 7331 (PART 1), 2012.
References
1. Cruz C., González J. R., Pelta D. Optimization in dynamic environments: a survey on problems, methods and measures // Soft Computing, 15 (7), Springer-Verlag, 2011.
2. Nguyena T. T., Yang S., Branke J. Evolutionary dynamic optimization: A survey of the state of the art // Swarm and Evolutionary Computation 6, 2012.
3. Weicker K. Evolutionary algorithms and dynamic optimization problems // Der Andere Verlag, 2003.
4. Ivanov I. A., Sopov E. A. Issledovaie effektivnosti samokonfiguriemogo koevolutsionnogo algoritma reshenia slozhnih zadach mnogokriterialnoi optimizatsii (On performance investigation of self-configured co-evolutionary algorithm for complex multi-objective optimization problem) // Control systems and information technologies, 1.1 (51), 2013.
5. Semenkin E., Semenkina M. Self-configuring genetic algorithm with modified uniform crossover operator // ADVANCES IN SWARM INTELLIGENCE (ICSI'2012). LNCS 7331 (PART 1), 2012.
© Сопов Е. А., 2014
МОДИФИКАЦИЯ АЛГОРИТМА SAWT C ВЕСОВЫМИ КОЭФФИЦИЕНТАМИ
А. В. Спирина
Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева Российская Федерация, 660014, г. Красноярск, просп. им. газ. «Красноярский рабочий», 31
E-mail: s_nastia@mail.ru
Рассматривается модификация алгоритма SAWT для вычисления расстояния Левенштейна, основная идея которой заключается в применении весовых коэффициентов для определения фонетической схожести слов.
Ключевые слова: алгоритмы нечеткого поиска, расстояние Левенштейна, строковые метрики.
