Diego Andina Universidad Polit´ecnica de Madrid, Spain Donald Davendra VSB - Technical University of Ostrava. Pavel Kord´ık Czech Technical University, Czech Republic Pavel Kr¨omer VSB-Technical University of Ostrava,.
Special Sessions
Vicente Martin Unibersidad ti Politekniko ti Madrid, Espania Vil ́em Nov ́ak Unibersidad ti Ostrava, Republika ti Tseko
Soft Computing Models for Control Theory & Applications in Electrical Engineering
Soft Computing Models for Biomedical Signals and Data Processing
Advanced Soft Computing Methods in Computer Vision and Data Processing
Organising Committee
Contents
Evolutionary Computation and Optimization
Intelligent Systems
Classification and Clustering Methods
Networks and Communication
Applications
A Hybrid Discrete Differential Evolution Algorithm for Economic Lot Scheduling Problem with Time Variant
Lot Sizing
1 Introduction
Metal forming and plastic production lines (press lines, and plastic and metal extrusion machines), where each product requires a different die to be set up on the relevant machinery. It explicitly handles the problems caused by setup times and always gives a feasible schedule as proved by Dobson [5].
2 ELSP Problem Formulation and Algorithm
The Proposed HDDE Algorithm
This lower bound is tighter than the one obtained by using the so-called independent solution, where each product is taken in isolation by calculating its economic production quantity. If the optimal cycle length for an item is Ti', the production frequency (fi) for that item can be obtained according to the following formula: '. ') max(.
Discrete Differential Evolution
The Cross-Over operator. The Cyclic Crossover Operator [12] is of immense importance in the proposed HDDE because this operation is carried out at several
The mutant individual is obtained by perturbing the previous generation's best solution in the target population. Therefore, we obtain the difference variances by applying a perturbation operator (mutations) to the best solution present in the previous generation of the target population.
3 Results and Discussion
Unlike its continuous counterpart, differential evolution is achieved by stochastic reshuffling of the most suitable solution of the previous generation in the target population. The value of mutationprob is set empirically by taking several runs of the proposed algorithm.
4 Conclusions and Further Work
Khouza, M., Michalewicz, Z., Wilmot, M.: Using Genetic Algorithms to Solve the Problem of Economic Lot Size Planning.
An Ordinal Regression Approach for the Unequal Area Facility Layout Problem
1 Introduction
This paper also provides a summary of some of the most commonly used and successful methods in ordinal. This paper is organized as follows: A brief analysis of the ordinal algorithms used is given in Section 2.
2 Methodology
Kernel Discriminant Learning for Ordinal Regression
Finally, in Section 4 we present the results of the analyzed algorithms compared with other state-of-the-art methods, and in Section 5 the conclusions and future work are summarized.
Ensemble for Ordinal Regression by Frank & Hall
Extended Binary Classification
It is considered to be one of the pioneer methods of ordinal regression since it was proposed in 1980 [21]. It uses a threshold method based on logistic regression and also assumes an assumption of stochastic ordering of the space X: for each pairx1andx2 it is convinced that P(y≤Ci|x1)≥P(y≤Ci|x2)orP(y≤Ci |x1) ≤P(y≤Ci|x2).
3 Experiments
- Dataset Description
- Methods Compared The methods compared are then
- Evaluated Measures
- Evaluation and Model Selection
Mean Absolute Error (MAE): it is defined as the average deviation in the absolute value of the predicted class from the true class. The results are taken as the mean and standard deviation of the measurements over the 30 test sets.
4 Results
In summary, we consider the KDLOR algorithm to be the most suitable for the proposed problem, as it achieved better performance in three of the four metrics considered and a smaller standard deviation (indicating that the results are more robust). This change to the KDLOR would change the model's thresholds according to the inverse of the number of patterns per class (so it would give more meaning to the minority classes).
5 Conclusion and Future Work
Babbar-Sebens, M., Minsker, B.S.: An interactive genetic algorithm with mixed initiative interaction for planning multi-criteria groundwater monitoring. Garc´ıa-Hern´andez, L., Salas-Morera, L., Arauzo-Azofra, A.: An Interactive Genetic Algorithm for the Object Allocation Problem in Unequal Areas.
A Soft Computing Approach to Knowledge Flow Synthesis and Optimization
2 Knowledge Flows
3 Related Evolutionary Computation Methods
Approaches to Evolution of Graphs
NEAT puts no limits on the size of the graph, but does not respect arities and I/O types of the actions. The most promising one seems to be the embryonic STGP proposed in [9], as it allows us to develop KFs of arbitrary size and can respect both I/O arities and types of actions.
Embryonic Strongly Typed Genetic Programming (STGP)
4 Evolution of Knowledge Flows
Our Approach
Based on problem analysis, we propose to modify embryonic STGP according to [9] to construct the KF based on a template. All actions and their parameters (both numerical and categorical) can be easily included in the genotype through the use of type-specific subtrees.
Cross-Validation Process Grammar
The design of the rest of the grammar is completely analogous and can easily be designed for different knowledge streams that solve different tasks. Very complex information can be encoded in trees, allowing us to move down the hierarchy and design actions that are appropriate for given layers of abstraction.
5 Experiments
Although the classification accuracy has only improved by percentage units during the evolution, in the field of DM, this is not insignificant. This is consistent with results obtained by a different strategy [14], where a complex set of models was also obtained for the vehicle data.
6 Conclusion
The Combination of Bisection Method and Artificial Bee Colony Algorithm
Artificial Bee Colony Algorithm
The number of employed bees is equal to the number of solutions in the population. Similar to the worker bee, she makes a change to the source position in her memory and checks the amount of nectar.
Bisection method algorithm )
Provided that the amount of nectar of the new one is higher than that of the previous source, the bee learns the position of the new source by heart and forgets the old one. Provided her nectar is higher than the previous one, the bee learns the new position by heart and forgets the old one.
2 The Combination of Bisection Method and Artificial Bee Colony Algorithm
The combination of bisection method and artificial bee colony algorithm 37 We illustrate this algorithm with some examples and compare results of proposed method with Newton method and ABC algorithm.
3 Numerical Examples
Example1
The result of proposed method and ABC algorithms to solve the fixed point problem in example.1.
Example2
2. The result of proposed method and ABC algorithms for solving the fixed point problem in example.2.
Example3
Newton's method loops and cannot obtain a good approximation of the solution. 3. The result of the proposed method and ABC algorithms for solving the fixed point problem of the example.3.
4 Conclusion
Karaboga, D., Basturk, B.: A powerful and efficient algorithm for numerical function optimization: Artificial Bee Colony (ABC) algorithm. Mansouri, P., Asady, B., Gupta, N.: A new iteration method for solving hard problems (nonlinear equations) with the artificial bee colony algorithm.
A System Learning User Preferences for Multiobjective Optimization
As a consequence, the participation of the designer is essential to incorporate quality considerations into the design. In this way, fatigue can be avoided and the search can be performed much faster, which is especially useful in the context of the large search space of object representations.
2 Structure of the Proposed System
This paper is organized as follows: a brief analysis of the built system is given in Section 2. Finally, we present some of the results obtained in Section 4, and Section 5 summarizes the conclusions and future work.
3 Learning Facility Layout Preferences from the Expert
Selection of the Algorithm
Learning User Preferences for MO Optimization of Facility Layouts 49 – Cost Matrix #1: Typical cost matrix for the nominal classifications, assuming. The cost matrix #4 is the one hybridized with the ordinal standard and the quadratic standard, and it obtains the optimal results.
4 Combining Subjective and Objective Criteria
Case Results
5 Conclusions and Future Work
This study investigates the bio-inspired artificial immune systems (AIS) as a pure metaheuristic soft computing solver of the LOP. A number of heuristics and soft computing algorithms have been used to solve LOP instances: greedy algorithms, local search algorithms, elite tabu searches, distributed searches and repetitive local searches [21,9,15]. The metaheuristic algorithms used to solve LOP in the past include genetic algorithms [12,13], differential evolution [23] and ant colony-based algorithms [4].
Genetic Programming of Augmenting Topologies for Hypercube-Based Indirect Encoding of Artificial
Neural Networks
The CPPN is then used to determine the synaptic weights of all possible connections found in the substrate (the coordinates of all pairs of neurons are passed to the inputs). Here we combine the advantages of both GP and NEAT and propose a new algorithm called Genetic Programming of Augmenting Topologies (GPAT).
2 Background
Genetic programming [8] is a well-known evolutionary approach which develops syntactic trees (or forests of trees). In this paper, we use a slightly simplified version of GP as abase algorithm, omitting a commonly used crossover operator, as it was not found beneficial by Buk [5] for Hypercube-based domain.
3 Our Approach
Apply the Cauchy mutation as described in the GP section to each tree constant with probability pCM. With probability pRC, for each tree constant, it resets its value to a random number from the given interval [−aR,aR].
4 Experimental Setup
The controller has a single output: for output values less than −0.5, the robot turns left and moves forward, for values greater than 0.5, it turns right and moves forward, otherwise it just takes a step forward. Correspondence is calculated as the number of all correct output bits in all test patterns (contrary to [11].) The problem is considered solved when all input patterns have produced correct output patterns.
5 Results
On the other hand, it performed worse than GP in all of them, with the exception of both maze tasks (significantly better) and Bit Shift (non-significantly better). However, for all indirect coding problems with the exception of visual discrimination, setting α=1 improves performance.
6 Conclusions
Finally, with a correct choice of the parameters K,α,andβ, GPAT is a significant winner on all problems, except the aforementioned 2D-K, 3D-K and 4D-V, dominated by GP and MAZE-2 and Parity where GPAT is not significantly worse than GPAT-R and GP. In: Proceedings of the 9th Annual Conference on Genetic and Evolutionary Computation - GECCO 2007, pp.
Master Slave LMPM Position Control Using Genetic Algorithms
The main idea of designing controller parameters using GA was publicly accepted in the 1990s, but remains popular in the present, which is proven by a number of articles in relevant journals [6][7]. These biologically motivated computing activities have waxed and waned over the years, but since the early 1980s they have all undergone a resurgence in the computing research community.
2 LMPM Position Control
Position Servo-Drive, Implementation Block Scheme
- Master-Slave Generator
- Precorrection Constants
The quality model of regulation system is highly sufficient, provided you are familiar with the scale of the most important parameters. In this paper, 4D master slave generator is used and it generates state variables for position, velocity, acceleration and jerk (Fig.3).
Luenberger Observer
3 Controller Design Methods
Pole Placement Method
Genetic Algorithm
First, a random population is generated with a predetermined number of chromosomes in one population within prescribed limits for the controller, lead compensator, and Luenberger observer parameters (see Table 1 for corresponding values). Then, the number of crossovers and mutations is applied to the population to achieve a higher chance of reaching the global optimum.
4 Simulation Results
5 Conclusion
Master Slave LMPM Position Control Using Genetic Algorithms 81 Second experiment represents linear servo-drive behavior using parameters obtained from GA and Pole Placement method with the parameters listed in Table 3 and Table 4 (Fig. 7] Wang, Y.P., Hur, D.R. , Chung , H.H. , Watson , N.R. , Arrillaga , J. , Matair , S.S. : A genetic algorithm approach to design an optimal PI controller for static VAr compensator.
An Approach of Genetic Algorithm to Model Supplier Assessment in Inbound Logistics
This paper presents a new model for supplier evaluation based on genetic algorithm that is used in a multinational company with operations in more than a dozen countries in the world as well as in Serbia. Supplier evaluation is an ongoing process and its results should be communicated to suppliers so that they can create a corrective action plan to develop and improve their weaknesses.
2 Literature Review
Literature Review of Soft Computing
Soft computing (SC) methods have been successfully applied to solve nonlinear problems in business, engineering, medicine, and science. Hybrid methods are based on combining procedures from different optimization methods to improve the accuracy and effectiveness of the base optimization methods.
Literature Review of Supply Chain
For example, in science, a new hybrid method based on fuzzy neural network learning algorithm for solving differential equation with fuzzy initial value is presented in [7]. Based on the aforementioned findings, it was found that price/cost is not the most adopted criterion.
3 The Previous Research
For each question regarding supplier performance, a target level of performance is set on a scale of 1 to 3 so that the sum of the entire performance set should be 50. It is possible to extend the previous performance supplier system with some other relevant issues that represent a new perceived supplier performance.
4 The GA Performance Value Constrained Model
5 Experimental Results
Company year Empirical surface GA value Distinction model without constraint constraint model GA & Empirical.
6 Conclusion and Future Work
The main purpose of supplier evaluation and selection processes is to ensure a successful and long-term collaboration between all parties in a supply chain. Ho, W., Xu, X., Dey, P.K.: Multi-criteria decision-making approaches for supplier evaluation and selection: a literature review.
Optimization of the Batch Reactor by Means of Chaos Driven Differential Evolution
2 Characteristics of the Batch Processes
3 Description of the Reactor
This technique allows in part to control the temperature of the reaction mixture by controlled feeding of the input chemical FK. The description of the reactor uses a system of four equilibrium equations (1) and one equation (2) representing the "k" term.
4 Differential Evolution
Batch Reactor Optimization Using Chaos-Driven Differential Evolution 95 Chemical FC (Filter Cake) flows into the reactor through the input referred to as “Chemical FC Input”, with parameters temperature-TFK, mass flow rate-mFC and specific heat cFC . For illustration, x4,2,0 represents the fourth value of the second solution in the first generation.
5 Chaotic Maps
Dissipative Standard Map
6 Problem Design
7 Results
Another fact that should not be neglected is the shortened duration of the process (about 20740 s compared to the original about 25000 s). Results of optimization, Simulation of the best solution, course of the weight of reaction mixture (top left), concentration of chemical FK (top right), temp.
8 Conclusions
8] Arpornwichanop, A., Kittisupakorn, P., Mujtaba, M.I.: Online dynamic optimization and control strategy for improving the performance of batch reactors. 17] Mujtaba, M., Aziz, N., Hussain, M.A.: Neural network based modeling and control in batch reactor. tive control based on genetic algorithms for the temperature control of a batch reactor.
Differential Evolution Classifier with Optimized Distance Measures for the Features in the Data Sets
In [10], however, we applied the same distance measure to all features in the dataset, while the proposed approach is optimizing the distance measures individually for each feature of the classified data. Clearly, an optimal distance measure for one feature may not be optimal for another feature in the same data set.
2 Differential Evolution Based Classifier with Optimized Distances for the Features in the Data Sets
Differential Evolution Based Classification
In addition, we also need to determine the possible parameters related to each distance measure and the class prototype vectors representing each class. Each population member,vi,G, as well as each new sample solution,ui,G, contains the class vectors for all classes and the power value p.
The Proposed Extension for Optimizing Distance Measures Individually for Each Feature in the Data Set
Once we have the distances between the samples and the class vectors, we can then make our classification decision according to the shortest distance. After dividing the vectorvi,Gis into corresponding parts, we can calculate the distances between the samples and the class vectors.
3 Classification Experiments and Comparisons of the Results
Also compared to BPNN, the proposed method gave significantly higher average accuracy in a confidence interval of 0.999. The results of the comparisons with the Horsecolic data show that the proposed method achieved a significantly higher average classification accuracy compared to the KNN, BPNN and DE classifications.
4 Discussion
Chauhan, N., Ravi, V., Chandra, DK: Differential evolution of a trained wavelet neural network: application to bank failure prediction. Triguero, I., Garcia, S., Herrera, F.: Differential evolution to optimize the positioning of prototypes in nearest neighbor classification.
Modifications of Differential Evolution with Composite Trial Vector Generation Strategies
One of these is the binomial crossover, which generates a new trial vector using the following rule. D}, andU1,U2, . . ,UDare independent random variables uniformly distributed in[0,1).CR∈[0,1]is a control parameter that affects the number of elements to be exchanged by the crossover.
2 DE with Composite Trial Vector Generation Strategies
The first is the number of function evaluations (nfe) required to complete the search. The second is the Reliability Rate (R) of the search, expressed as the number of runs.
5 Conclusion
A survey of the state-of-the-art. Mallipeddi, R., Suganthan, P.N., Pan, Q.K., Tasgetiren, M.F.: Differential Evolution Algorithm with Ensemble of Parameters and Mutation Strategies. IEEE Transactions on Evolutionary Computation 13, 398–. Storn, R., Price, K.V.: Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces.
Multi-Objective Differential Evolution on the GPU with C-CUDA
4,5] proposed a multi-objective approach based on DE, called Multi-Objective Differential Evolution (MODE) by the authors, which builds on the ideas from NSGAII. Given this scenario, this work studies the implementation of Multi-Objective Differential Evolution on the GPU with C-CUDA (MODE+GPU), evaluating the gain in processing time compared to the sequential version.
2 Multi-objective Differential Evolution on the GPU
- Benchmark Functions
- Computational Environment
- Parameters
- Results
The next section describes the experimental setups and presents the performance comparison between the algorithms. The probability in the second experiment is also based on random, but did not influence the results.
4 Conclusions
Batista, L.S., Campelo, F., Guimar˜aes, F.G., Ram´ırez, J.A.: Pareto Coneε dominance: improving convergence and diversity in multi-objective evolutionary algorithms.
Competitive Differential Evolution Algorithm in Comparison with Other Adaptive Variants
2 Differential Evolution
The experimental vectory is created by crossing two parental vectors, the current (target) vectorjaxi and the mutant vectorjav. The vector r1=xi is randomly selected from P, r2 is randomly selected from the union PA of the current population P and the archive A. The crossover operator constructs a trial vectory from the current individual axi and the mutant vectorjav.
3 Competitive Differential Evolution
4 Experiments and Results
The order of the algorithms is slightly different from the comparison of the entire set of functions. The average rank of both b6e6rl modifications is in the middle of the algorithms in the experimental comparison.
Differential Evolution and Perceptron Decision Trees for Fault Detection in Power Transformers
Dissolved Gas Analysis in Power Transformers
The problem is to diagnose the fault based on the data describing the gases generated under those conditions. Thus, the problem of analyzing dissolved gases consists of finding rules that can identify faults in the transformers.
Differential Evolution
In those nodes there is information about the class that should correctly classify the data. With the equation wx+θ =z, where w are weight scalars, x are the data features and θ is a constant, we can linearly divide the search space into areas where z>0 and areas where z<0.
3 Methodology
Operators
As for leaf nodes, a discrete solution intersection approach should be used [17]. Differential Evolution and Perceptron Decision Trees for Fault Detection 147 new individual survival operator was developed to always keep new candidates in the population.
Legitimacy
In this new test, we performed 15 independent executions of the algorithm for each database with 2000 generations. The figures show the minimum, average and maximum error for the 15 runs in the validation set.
5 Discussion
All parameters of the algorithm were adjusted separately, as there is no interaction between the factors. Another important point in defining the objective function is defining the costs involved in the process.
Multiobjective Pareto Ordinal Classification for Predictive Microbiology
Ordinal Model
The major problem in ordinal classification is that there is no idea of the exact distance between classes. A function f(x) that predicts the actual value outcomes and tries to discover the nature of the hypothesized underlying outcome.
Metrics and Methodology
3 Description Problem
4 Experiments
Experimental Design
The first strategy selects the best model in the AMAE, which is the upper individual from the Pareto front. The second strategy selects the best model in the MMAE, which is the bottom individual with the Pareto front.
Comparison Methods
Once the Pareto front is built into the latest generation of training, two selection strategies are used to choose the individuals. Since the MPENSGA2 algorithm is stochastic, the algorithm was run 30 times and the mean and standard deviation were obtained from the 30 individuals for the upper and lower extremes.
Results
The best models are generated by the MPENSGA2-A method, this method achieves the best results in all measurements. Thus, we propose the MPENSGA2 algorithm to solve the two problems of predictive microbiology, specifically the top best model of the Pareto front which maximizes the value of AMAE.
5 Conclusions
FNTs and FRs have been used to estimate the time error parameter in a real dental grinding process. The parameters used for the evolution of FNT and FR are shown in Table 2.
Self-organizing Migration Algorithm on GPU with CUDA
Self-organizing migration algorithm (SOMA) was chosen for a subset of its features that are not common in other EAs and make it suitable for parallel architecture with limited communication between processing units. Next part describes methodology used in testing the algorithms performance followed by results of the tests and conclusion.
2 Methods
SOMA
Self-organizing migration algorithm on GPU with CUDA 175, where ML is the number of current migration rounds, xMLi,j,start is the position of the active person at the beginning of the current migration, xMLL,j is the position of the leader, t ∈ [0 ;. At the end of the migration (individual made all steps lower than PathLength), the individual is set to a position that had the best fitness value during the current migration.
The population size was set to 3,000, the dimension to 50 and the number of migrations to 100, the number of threads per block was changed to powers of 2. All values above zero mean that the shared manager version was faster for a given number of threads per block.
3 Results
First test was aimed at how well cuSOMA scales to increase in population size. Second test used constant population size and the dimension of the test functions was increasing.
4 Conclusion
IEEE Congress Evolutionary Computation CEC 2009, pp. de Veronese, L.P., Krohling, R.A.: Swarm's flight: Acceleration of the particles using C- CUDA. Senkerik, R., Zelinka, I., Oplatkova, Z.: Comparison of differential evolution and SOMA in the task of chaos control optimization - Extended study.
Urban Traffic Flow Forecasting Using Neural-Statistic Hybrid Modeling