查看更多>>摘要：Value-based reinforcement learning (RL) algorithms have been widely applied in traffic signal studies. There are, however, several problems in jointly controlling traffic lights for a large transportation network. First, the discrete action space exponentially explodes as the number of intersections to be jointly controlled increases. With its model structure, the original deep Q-network (DQN) could not accommodate a large action space. The problem was resolved by revising the output structure of a DQN holding the framework of a single-agent RL algorithm Second, when mapping traffic states into an action value, it is difficult to consider spatio-temporal correlations over a large transportation network. A deep graph Q-network (DGQN) was devised to efficiently accommodate spatio-temporal dependencies on a large scale. Finally, training the proposed DGQN with a large number of joint actions requires much time to converge. An asynchronous update methodology with multiple actor learners was devised for a DGQN to quickly reach an optimal policy. By combining these three remedies, a DGQN succeeded in jointly controlling the traffic lights in a large transportation network in Seoul. This approach outperformed other "state-of-the-art "RL algorithms as well as an actual fixed-signal operation. The proposed DGQN decreased the average delay of the current fixed operation to 55.7%, whereas those of reference models DQN-OGCN and DQN-FC were 72.5 and 92.0%, respectively. (c) 2022 Elsevier B.V. All rights reserved.
Jurczuk, KrzysztofCzajkowski, MarcinKretowski, Marek
19页
查看更多>>摘要：Evolutionary algorithms (EAs) are naturally prone to parallel processing. However, when they are applied to data mining, the fitness calculations start to dominate and the typical population-based decomposition limits the parallel efficiency. When dealing with large-scale data, the scalable solution may become a real challenge. In this article, we propose a GPU-based parallelization of evolutionary induction of model trees. Such trees are a special case of decision tree (DT) that is designed to solve regression problems. The evolutionary approach allows not only a robust prediction but also to preserve the simplicity of DTs. However, the global approach is much more computationally demanding than state-of-the-art greedy inducers, and thus hard to apply to large-scale data mining directly. A parallelized induction of model trees (with univariate tests in the internal nodes and multiple linear regression models in the leaves) requires a carefully designed decomposition strategy. Six GPUsupported procedures are designed to successively: redistribute, sort and rearrange dataset samples, next, calculate models and fitness, and finally gather the results. Experimental validation is performed on real-life and artificial datasets, using various (low- and high-end) GPU accelerators. Results show that the GPU-supported solution enables time-efficient global induction of model trees on large-scale data, which until now was reserved for greedy methods. The obtained speedup is very satisfactory (even up to hundreds of times). The solution is scalable for datasets of different sizes and dimensions. (c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：In multi/many-objective optimization, a decision maker (DM) may often be interested in examining only a small set of solutions instead of the entire Pareto optimal front (PF). Such solutions are referred to as solutions of interest (SOI) in some recent studies. A number of methods have been proposed to identify SOIs in an offline or online setting using measures based on reflex angle, bend angle, expected marginal utility, etc. However, these measures only account for the desirable trade-offs in the objective space. On the other hand, the variable space information is often critical in practical scenarios as it relates directly to the implemented design. For example, a DM may additionally require that the obtained solutions are robust, i.e., insensitive to variable perturbations, or look significantly different in the variable space, thereby offering multiple equivalent designs to achieve similar performance. These require formulation of new measures and search strategies that simultaneously consider both objective and variable spaces while identifying SOIs. In this paper, we develop an approach that can identify a given number of SOIs for DM's consideration for three different scenarios: (a) purely based on objective space, (b) simultaneous consideration of objectives and robustness, and (c) simultaneous considerations of objectives and equivalent designs. Towards this end, we first define the relevant quantitative measures and illustrate their use for offline selection for a few 2-3 objective test problems. Thereafter, we design an online algorithm that can identify the SOIs and bias the search towards the SOIs based on the scenarios listed above. Lastly, we also present results on two practical examples: a 2-objective welded beam and a 5-objective wind-turbine design problem. (C) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：Classification accuracy and interpretability are crucial importance for recognizing seizures based on electroencephalogram (EEG) signals. This study presents a novel deep ladder-type Takagi-Sugeno-Kang (TSK) fuzzy classifier (D-LT-TSK) that alternately utilizes horizontal progressive learning and longitudinal leapfrogging learning styles. Based on the nonuniform probability distribution co-generated by the distance correlation (DC) coefficient and random bias matrix, a feature activity adjustment mechanism (DC-FAM) is adopted to adjust the activity of each feature to realize the evolution from full connection to partial connection between the input layer and rule layers of the TSK classifier. Feedforward and feedback neural networks are combined to learn consequent parameters in the Then-part of fuzzy rules, for the sake of strengthening the approximation performance and achieving fast converge capability. To take full advantage of valuable decision-making information, D-LT-TSK is learned in the horizontal progressive and longitudinal leapfrogging learning style by mapping the decision-making information of learning modules into the original input space. Experimental results demonstrated that (1) the highly interpretable D-LT-TSK be capable of yielding satisfactory classification performance by utilizing short fuzzy rules, and (2) the optimization algorithm in the Then-part enhanced the approximation performance and accelerate the convergence speed. (c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：Due to the absence of any specialized drugs, the novel coronavirus disease 2019 or COVID-19 is one of the biggest threats to mankind Although the RT-PCR test is the gold standard to confirm the presence of this virus, some radiological investigations find some important features from the CT scans of the chest region, which are helpful to identify the suspected COVID-19 patients. This article proposes a novel fuzzy superpixel-based unsupervised clustering approach that can be useful to automatically process the CT scan images without any manual annotation and helpful in the easy interpretation. The proposed approach is based on artificial cell swarm optimization and will be known as the SUFACSO (SUperpixel based Fuzzy Artificial Cell Swarm Optimization) and implemented in the Matlab environment. The proposed approach uses a novel superpixel computation method which is helpful to effectively represent the pixel intensity information which is beneficial for the optimization process. Superpixels are further clustered using the proposed fuzzy artificial cell swarm optimization approach. So, a twofold contribution can be observed in this work which is helpful to quickly diagnose the patients in an unsupervised manner so that, the suspected persons can be isolated at an early phase to combat the spread of the COVID-19 virus and it is the major clinical impact of this work. Both qualitative and quantitative experimental results show the effectiveness of the proposed approach and also establish it as an effective computer-aided tool to fight against the COVID-19 virus. Four well-known cluster validity measures Davies-Bouldin, Dunn, Xie-Beni, and beta index are used to quantify the segmented results and it is observed that the proposed approach not only performs well but also outperforms some of the standard approaches. On average, the proposed approach achieves 1.709792, 1.473037, 1.752433, 1.709912 values of the Xie-Beni index for 3, 5,7, and 9 clusters respectively and these values are significantly lesser compared to the other state-of-the-art approaches. The general direction of this research is worthwhile pursuing leading, eventually, to a contribution to the community. (c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：As a powerful optimization technique, multi-objective particle swarm optimization algorithms have been widely used in various fields. However, performing well in terms of convergence and diversity simultaneously is still a challenging task for most existing algorithms. In this paper, a multi-objective particle swarm optimization algorithm based on two-archive mechanism (MOPSO_TA) is proposed for the above challenge. First, two archives, including convergence archive (CA) and diversity archive (DA) are designed to emphasize convergence and diversity separately. On one hand, particles are updated by indicator-based scheme to provide selection pressure toward the optimal direction in CA. On the other hand, shift-based density estimation and similarity measure are adopted to preserve diverse candidate solutions in DA. Second, the genetic operators are conducted on particles from CA and DA to further enhance the quality of solutions as global leaders. Then the search ability of MOPSO_TA can be improved by performing hybrid operators. Furthermore, to balance global exploration and local exploitation of MOPSO_TA, a flight parameters adjustment mechanism is developed based on the evolutionary information. Finally, the proposed algorithm is compared experimentally with several representative multi-objective optimization algorithms on 21 benchmark functions. The experimental results demonstrate the competitiveness and effectiveness of the proposed method.(c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：Occupancy detection and prediction are two well-established problems which can be improved further to achieve higher accuracy in both cases than the existing solutions. To achieve the desired higher accuracy, proposed OccupancySense model detects human presence and predicts indoor occupancy count by the fusion of Internet of Things (IoT) based indoor air quality (IAQ) data along with static and dynamic context data which is a unique approach in this domain. This data fusion helps us to achieve higher forecasting accuracy along with the integration of state of the art gradient boosting based categorical features supported CatBoost algorithm. For comparison, other commonly used machine learning classification and regression algorithms, e.g., Multiple Linear Regression (MLR), Decision Tree (DT), Random Forests (RF) and Support Vector Machine (SVM) for regression and Logistic Regression (LR), Naive Bayes (NB), Decision Tree (DT) and Random Forest (RF), Support Vector Machine (SVM) for classification, were also assessed during this experiment. Out of these, CatBoost outperformed other models when considered in terms of accuracy. Hence, CatBoost is used as the core of the OccupancySense design and we have validated the proposed model by a real-world case study with continuous 91 days of indoor data, having 33 unique external features. These features are collected directly as well as derived from the collected data. To handle these features, feature engineering plays a key role in the OccupancySense model. The speciality of this model is, it is non-intrusive one but have high predictive power. It can detect occupancy and predicts headcount along with occupancy density of the room pretty accurately with 99.85%, 93.2% and 95.6% respectively (with 10 fold cross-validation) which outperforms other state of the art models. (c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：Moth-Flame Optimization (MFO) algorithm, which is inspired by the navigation method of moths, is a nature-inspired optimization algorithm. The MFO is easy to implement and has been used to solve many real-world optimization problems. However, the MFO cannot balance exploration and exploitation well, and the information exchange between individuals is limited, especially in solving some complex numerical problems. To overcome these disadvantages of the MFO in solving the numerical optimization problems, a covariance-based Moth-Flame Optimization algorithm with Cauchy mutation (CCMFO) is proposed in this paper. In the CCMFO, the concept of covariance is used to transform the individuals of the moths and flames from the original space to the eigenspace and update the positions of moths, which can better improve the information exchange ability of the flames and moths in the eigenspace. In addition, Cauchy mutation is utilized to improve the exploration. And the CCMFO is compared with the other 22 algorithms on CEC 2020 test suite. The test results show that the CCMFO is better than other population-based optimization algorithms and MFO variants in search performance, while its performance is statistically similar to CEC competition algorithms. Furthermore, the CCMFO is compared with the other 12 algorithms on CEC 2020 real -world constrained optimization problems, and the results show that the CCMFO can effectively solve real-world constrained optimization problems. Finally, the CCMFO is used to optimize the tracking controller parameters of continuous casting mold vibration displacement. The experimental results based on the experimental platform show that the CCMFO can effectively reduce the difficulty of parameter selection and improve the tracking accuracy. (c) 2022 Elsevier B.V. All rights reserved.
查看更多>>摘要：This work proposes a novel approach to evaluate and analyze the behavior of multi-population parallel genetic algorithms (PGAs) when running on a cluster of multi-core processors. In particular, we deeply study their numerical and computational behavior by proposing a mathematical model representing the observed performance curves. In them, we discuss the emerging mathematical descriptions of PGA performance instead of, e.g., individual isolated results subject to visual inspection, for a better understanding of the effects of the number of cores used (scalability), their migration policy (the migration gap, in this paper), and the features of the solved problem (type of encoding and problem size). The conclusions based on the real figures and the numerical models fitting them represent a fresh way of understanding their speed-up, running time, and numerical effort, allowing a comparison based on a few meaningful numeric parameters. This represents a set of conclusions beyond the usual textual lessons found in past works on PGAs. It can be used as an estimation tool for the future performance of the algorithms and a way of finding out the upper limit of the performance if the number of used cores increases.(C) 2022 Elsevier B.V. All rights reserved.
Vijayan, S. Venkata K.Mohanta, HarePani, Ajaya Kumar
15页
查看更多>>摘要：Real time estimation of target quality variables using soft sensor relevant to time varying process conditions will be a significant step forward in effective implementation of Industry 4.0. Generalized Regression neural network (GRNN) has been used as a steady state quality monitoring soft sensor with reasonable estimation accuracy. However, the accurate prediction capability of GRNN has rarely been explored in a time varying environment. This article reports design of adaptive soft sensor using GRNN as a local model in Just-in-Time learning (JITL-GRNN) framework. The JITL-GRNN adaptive soft sensing technique is further investigated in various dimensions such as, the effect of different similarity index criteria and relevant dataset size on model prediction accuracy and model computation time. Performance of the proposed JITL-GRNN soft sensor is investigated by assessing its prediction accuracy on two benchmark industrial datasets. In addition, dynamic Non-linear autoregressive with exogenous inputs (NARX) neural network model is also developed and the performance of NARX model was compared with the proposed JITL-GRNN model. Results show that the JITL-GRNN adaptive soft sensor has at par or better prediction capability than the NARX model and many other models reported in literature. (c) 2022 Elsevier B.V. All rights reserved.
NSTL
Elsevier