Many image features for image recognition have been proposed in previous studies. The dimensionality of the previously-proposed image features is mostly over several thousands. The higher the dimensionality of features, the higher the accuracy of image classification and object recognition. However, the use of these high dimensional features involves high computational costs and the difficulty of analysis of recognition processing. In this paper, we propose a hierarchical feature dimension reduction method which is based on cartesian genetic programming (CGP). The proposed method generates a predefined number of new features one at a time using CGP per layer. A CGP generates a new feature by combining the previously-proposed image features. The proposed method finally generates three-dimensional features to visualize the data landscapes. We performed experiments on the Graz dataset, the capsule endoscopy images and the INRIA person dataset, and with regard to the support vector machine, the decision tree and k-nearest neighbor classifiers, the classification results obtained using the features generated by the proposed method were better than those using the previously-proposed features. The visualization results of the three-dimensional features generated by the proposed method showed that the object images are broadly separated from the non-object images. Moreover, the proposed method reduced the computational costs of feature extraction and classification by reducing the dimensionality of features.
This paper presents a parameter tuning study of Differential Evolution (DE) algorithms, including standard DE as well as various adaptive DE algorithms (jDE, JADE and SHADE) for different computational budget scenarios. While there have been numerous parameter studies of DE for cheap computational budget scenarios where DE can utilize a relatively-large computational budget, there have been no previous studies of DE for expensive computational budget scenarios where only a small computational budget can be used for the search. Using the algorithm configuration tool SMAC, the DE variants are tuned independently for three different maximum number of evaluations: 102×D, 103×D and 104×D evaluations, where D is the benchmark problem dimensionality. Each of these tuned parameter settings is then tested in each of the three different budget scenarios, which enables us to analyze the effect of both the tuning and testing phase on the performance of the tuned algorithm. For the parameter tuning phase, we use the CEC2014 benchmarks as training problems, and for the testing phase, we use all 24 problems from the BBOB noiseless benchmark set. The experimental results show: (1) the parameter settings obtained by SMAC vary significantly depending on the maximum number of evaluations on the training phase, (2) DE algorithms perform well when the computational budget on the training phase and the testing phase are similar, (3) the standard DE algorithm performs better than some adaptive DE algorithms for low-dimensional problems in expensive scenarios.
Multi-Objective Genetic Algorithm (MOGA) is an application of Genetic Algorithm for solving Multi-objective Optimization Problems (MOPs). Many researches on MOGA have been actively reported. Generally, it is difficult to obtain the optimized solution satisfying all objective functions because of their trade-offs. Then, it is necessary to obtain Pareto solutions which are not inferior to other solutions in at least one objective function. In the application of MOGA to engineering design fields, it is not the goal to obtain high-performance Pareto solutions, because it is also important to analyze the obtained Pareto solutions and extract the knowledge in the problem. In order to analyze Pareto solutions obtained by MOGA, it is required to consider both the objective space and the design variable space. In this paper, we define Non-Correspondence in Linear Relationship between the objective space and the design variable space. We also try to extract the Non-Correspondence area in Linear Relationship with the index defined in this paper. This paper applies the proposed method to the trajectory designing optimization problem and extracts Non-Correspondence area in Linear Relationship in the acquired Pareto solutions. This paper also applied the index to the method of fitness approximation and showed that the reduction of the number of evaluation can be carried out.
The XCS classifier system is designed to evolve accurately generalized classifiers as an optimal solution to a problem. All classifiers are identified as either accurate or inaccurate on the basis of a pre-defined parameter called an accuracy criterion. Previous results suggested a standard setting of the accuracy criterion robustly performs on multiple simple problems so XCS evolves the optimal solution. However, there lacks a guideline of reasonable setting of accuracy criterion. This causes a problem that the accuracy criterion should be empirically customized for each complex problems especially noisy problems which is a main focus of this paper. This paper proposes a self-adaptation technique for the accuracy criterion which attempts to enable XCS to evolve the optimal solution on the noisy problems. In XCS-SAC(XCS with Self-Adaptive accuracy criterion), each classifier has its own accuracy criterion in order to find an adequate setting of accuracy criterion for each niche. Then, each classifier's accuracy criterion is updated with the variance of reward which its classifier has received. We test XCS-SAC on a benchmark classification problem (i.e., the multiplexer problem) with noise (the Gaussian noise and alternative noise). Experimental results show XCS-SAC successfully solves the noisy multiplexer problems as well as XCS but evolves a more compact solution including an optimal solution than XCS.
This paper presents a new relocation method for Adaptive Weighted Aggregation with Step Size Control Weight Adaptation (AWA-SSCWA) that is a powerful multi-start framework of scalarized decent methods for multi-objective continuous function optimization. AWA-SSCWA repeats two procedures: subdivision and relocation. Subdivision decides initial weight vectors and solutions for relocation. Relocation iteratively adapts weight vectors by repeating two procedures of optimization and weight adaptation in order to improve the coverage of an approximate solution set. Weight adaptation uses neighborhood relationship between weight vectors and solutions to estimate appropriate weight vectors which achieve an approximate solution set with good coverage. AWA-SSCWA employs lattice points, called addresses, to quickly calculate the neighborhood relationship between weight vectors and solutions. AWA-SSCWA has been reported to find good approximate solution sets in terms of coverage. However, AWA-SSCWA has a serious problem in that the performance of AWA-SSCWA deteriorates in terms of coverage when the neighborhood relationship of addresses does not match that of solutions. In order to remedy the problem, we propose a new version of AWA-SSCWA, named AWA with Enhanced Relocation (AWA-ER). In order to investigate the effectiveness of AWA-ER, we compared the performance of AWA-ER with that of AWA-SSCWA on three to five objective benchmark problems with twenty variables. As the result, we confirmed that AWA-ER outperformed AWA-SSCWA on the benchmark problems.
In the field of air conditioning control for office, improving office worker's thermal comfort with lower energy is required. Usually, single-objective mathematical optimization techniques are used for air conditioning control considering thermal comfort and energy saving. In such techniques, scalarization method (e.g. linearly weighting addition, constraint transformation) is used. However, it is difficult to match dynamic changes of temperature and/or consumed power requirements in such methods. In this study, we apply an evolutionary multi-objective optimization (EMO) method to decide optimal operation of an air conditioning system considering both thermal comfort and power consumption. We formulate two objective functions, thermal comfort and power consumption. Predicted Mean Vote (PMV) was used as the first objective function. PMV is widely known thermal comfort index of indoor conditions adopted as an ISO standard. An estimation method of power consumption of air conditioning from a heat load was used as the second objective function. The heat load is amount of heat provided or removed by an air conditioning system. In this study, we used OMOPSO (Optimized Multi-Objective Particle Swarm Optimization), one of the MOPSO algorithms, which uses an external archive and mutation operators. An archive of OMOPSO has a concept of ε-dominance, so in our study we varied parameter ε to finding better solutions. Comparison with constant temperature control, our results demonstrate the EMO method is effective to get a set of solutions for the best considering trade-off between thermal comfort and power consumption. By not using ε-dominance, we got wide range of PMV value recommended by ISO standard. After searching pareto solutions, if we can select an effectual solution adopting to requirements, optimal air conditioner control considering requirement is realizable.