evolutionaryalgorithms (EAs) show good performance in solving multi-objective optimization problems (MOPs). An EA needs to perform a substantial number of fitness evaluations. For the MOP with high complexity, the fi...
详细信息
evolutionaryalgorithms (EAs) show good performance in solving multi-objective optimization problems (MOPs). An EA needs to perform a substantial number of fitness evaluations. For the MOP with high complexity, the fit-ness evaluation functions are computationally expensive, making the evolutionaryalgorithms time-consuming. surrogate-assisted evolutionary algorithms (SAEAs) that apply surrogate models instead of fitness exact evalua-tion functions have successfully reduced the computational complexity of fitness evaluations. However, because training a surrogate model requires a certain amount of calculation, a large amount of calculation is required by the SAEA to train multiple surrogate models. Furthermore, most existing surrogate models may not achieve desired evaluation accuracy when processing medium-dimensional and high-dimensional MOPs. This paper pro -poses a novel surrogate model. The surrogate model can be applied in multi-objective optimization evolutionary algorithm based on decomposition (MOEA/D), which is a classic decomposition-based multi-objective optimiza-tion algorithm. The surrogate model is designed based on the convolutional neural network structure, and it is called the multi-objective parallel fitness evaluation network (MPFEN). An MPFEN model contains multiple sub-networks which can be applied as the surrogate models. By training the MPFEN model, we can obtain all sur-rogate models required by a MOEA/D simultaneously without training each required surrogate model separately. Therefore, the amount of calculation of training surrogate models in a MOEA/D is reduced. The evaluation accu-racy of the MPFEN model is tested by experiments. The experimental results show that the evaluation accuracy of MPFEN model is higher than that of other classical surrogate models in most cases. By applying the MPFEN model, the solution quality of SAEA is also improved.
Training surrogate models with high quality often requires a sufficient quantity of labelled data with a balanced distribution. However, obtaining enough labelled solutions for expensive optimization problems is chall...
详细信息
ISBN:
(纸本)9789819771806;9789819771813
Training surrogate models with high quality often requires a sufficient quantity of labelled data with a balanced distribution. However, obtaining enough labelled solutions for expensive optimization problems is challenging, let alone achieving a uniformly distributed training dataset. In this paper, we propose an expensive multi-objective evolutionary algorithm based on regional density ratio (MOEA-RDR) for solving computationally expensive problems. In MOEA-RDR, a new evaluation metric, integrating the uncertainty measures of Gaussian process models with the underlying assumptions of semi-supervised techniques, is introduced to select unlabelled solutions to participate in the training of surrogate models. A number of experiments are conducted on WFG test problems, and the experimental results show that our proposed method is more efficient than four state-of-the-art algorithms for solving computationally expensive multi-objective problems.
surrogate-assisted evolutionary algorithms (SAEAs) aim to use efficient computational models with the goal of approximating the fitness function in evolutionary computation systems. This area of research has received ...
详细信息
ISBN:
(纸本)9798400701207
surrogate-assisted evolutionary algorithms (SAEAs) aim to use efficient computational models with the goal of approximating the fitness function in evolutionary computation systems. This area of research has received significant attention from the specialised research community in different areas, for example, single and many objective optimisation or dynamic and stationary optimisation problems. An emergent and exciting area that has received little attention from the SAEAs community is in neuroevolution. This refers to the use of evolutionaryalgorithms in the automatic configuration of artificial neural network (ANN) architectures, hyperparameters and/or the training of ANNs. However, ANNs suffer from two major issues: (a) the use of highly-intense computational power for their correct training, and (b) the highly specialised human expertise required to correctly configure ANNs necessary to get a well-performing network. This work aims to fill this important research gap in SAEAs in neuroevolution by addressing these two issues. We demonstrate how one can use a Kriging Partial Least Squares method in place of the well-known Kriging method, which normally cannot be used in neuroevolution due to the high dimensionality of the data.
Novelty search's ability to efficiently explore the fitness space is gaining attention. Different novelty metrics, however, produce different search results. Here we show that novelty metrics are complementary and...
详细信息
ISBN:
(纸本)9798400701207
Novelty search's ability to efficiently explore the fitness space is gaining attention. Different novelty metrics, however, produce different search results. Here we show that novelty metrics are complementary and a multi-novelty approach improves the performance substantially. Specifically, we propose a multi-novelty search multi/many-objective algorithm (Curious II) that has both Euclidian distance and prediction-error novelty metrics. On the one hand, the Euclidian distance based novelty metric makes the sub-population explore subspaces with low crowd density and avoids premature convergence. On the other hand, the prediction-error novelty metric guides a subpopulation to explore subspaces with unexpected objective fitness. Experiments reveal that using both novelty metrics in a multi-novelty algorithm has strong benefits. Curious II was compared with two state-of-the-art algorithms and two novelty search-based algorithms on the WFG 1-8 test problem with up to 10 objectives. It outperforms all the others in 28 out of 32 tasks for the HV index and in 27 out of 32 tasks for the IGD index.
Deep Neural Networks (DNNs) have created a breakthrough in medical image analysis in recent years. Because clinical applications of automated medical analysis are required to be reliable, robust and accurate, it is ne...
详细信息
ISBN:
(数字)9781665467087
ISBN:
(纸本)9781665467087
Deep Neural Networks (DNNs) have created a breakthrough in medical image analysis in recent years. Because clinical applications of automated medical analysis are required to be reliable, robust and accurate, it is necessary to devise effective DNNs based models for medical applications. In this paper, we propose an ensemble framework of DNNs for the problem of medical image segmentation with a note that combining multiple models can obtain better results compared to each constituent one. We introduce an effective combining strategy for individual segmentation models based on swarm intelligence, which is a family of optimization algorithms inspired by biological processes. The problem of expensive computational time of the optimizer during the objective function evaluation is relieved by using a surrogate-based method. We train a surrogate on the objective function information of some populations and then use it to predict the objective values of each candidate in the subsequent populations. Experiments run on a number of public datasets indicate that our framework achieves competitive results within reasonable computation time.
Environmental selection of multiobjective evolutionaryalgorithms (MOEAs) is a key component that chooses promising solutions from a candidate set for later usage. Most environmental selection strategies choose soluti...
详细信息
Environmental selection of multiobjective evolutionaryalgorithms (MOEAs) is a key component that chooses promising solutions from a candidate set for later usage. Most environmental selection strategies choose solutions based on their function values. However, in real-world optimization problems, function evaluations can be time consuming. The necessity of a large number of function evaluations leads to the low efficiency of MOEAs. How to decrease the number of function evaluations is one of the main issues of MOEAs. The environmental selection can be regarded as a classification process. The selected solutions are the promising class, and the discarded solutions are the unpromising class. Based on this consideration, we propose a classification-assisted environmental selection (CAES) strategy in this paper to decrease the number of function evaluations in MOEAs. In the proposed method, solutions are divided into two classes. One is non-dominated solutions (i.e. promising class) and the other is dominated solutions (i.e. unpromising class). The classifier is built to classify the offspring solutions into these two classes. Only promising offspring are evaluated (unpromising ones are removed with no function evaluations). Therefore, the number of function evaluations is reduced. We integrate the proposed CAES strategy into six MOEAs. The effectiveness of the proposed CAES strategy is examined through computational experiments on various test suites and three real-world application problems. Our experimental results show that the proposed CAES strategy clearly reduces the number of function evaluations without severely degrading the search ability of the original MOEAs.
The key issue in handling multimodal multi -objective optimization problems (MMOPs) is to find multiple Pareto sets (PSs) corresponding to one Pareto front (PF). Therefore, learning the PSs is critical to facilitate s...
详细信息
The key issue in handling multimodal multi -objective optimization problems (MMOPs) is to find multiple Pareto sets (PSs) corresponding to one Pareto front (PF). Therefore, learning the PSs is critical to facilitate solving MMOPs while unfortunately, current research only focuses on PF learning which is helpless in finding multiple PSs by the information of one PF. Moreover, since the PSs of an MMOP are usually non-functional, traditional approximative function model -based PF learning is inapplicable. Consequently, developing new PS learning techniques is desired. Inspired by data -driven evolutionaryalgorithms, data can be used to train surrogate models to assist the algorithm. This article proposes an online data -driven PSs learning technique that aims to learn the topologies of PSs through a surrogate model to facilitate the search for PSs. Specifically, the Growing Neural Gas network is trained using non -dominated solutions to learn the topologies of PSs during the evolutionary process. Then, the nodes of the network are used to generate new solutions and adopted as reference points for environmental selection. A new algorithm is developed based on the PS learning technique for MMOPs. Experimental studies on three benchmark test suites and two different real -world applications demonstrate the superiority of our method over six state-of-the-art algorithms dedicated to MMOPs. The PSs learning technique can obtain the topologies of PSs and facilitate the search for them.
Real-world problems are often affected by uncertainties of different types and from multiple sources. algorithms created for expensive optimisation, such as model-based optimisers, introduce additional errors. We argu...
详细信息
Real-world problems are often affected by uncertainties of different types and from multiple sources. algorithms created for expensive optimisation, such as model-based optimisers, introduce additional errors. We argue that these uncertainties should be accounted for during the optimisation process. We thus introduce a benchmark as well as a new surrogate-assistedevolutionary algorithm to investigate this hypothesis further. The benchmark includes two function suites based on procedural content generation for games, which is a common problem observed in games research and also mirrors several types of uncertainties in the real-world. We find that observing and handling the uncertainty present in the problem can improve the optimiser, and also provides valuable insight into the function characteristics.
暂无评论