Advances in computational intelligence最新文献

Due to the outbreak of COVID-19 disease globally, countries around the world are facing shortages of resources (i.e. testing kits, medicine). A quick diagnosis of COVID-19 and isolating patients are crucial in curbing the pandemic, especially in rural areas. This is because the disease is highly contagious and can spread easily. To assist doctors, several studies have proposed an initial detection of COVID-19 cases using radiological images. In this paper, we propose an alternative method for analyzing chest X-ray images to provide an efficient and accurate diagnosis of COVID-19 which can run on edge devices. The approach acts as an enabler for the deep learning model to be deployed in practical application. Here, the convolutional neural network models which are fine-tuned to predict COVID-19 and pneumonia infection from chest X-ray images are developed by adopting transfer learning techniques. The developed model yielded an accuracy of 98.13%, sensitivity of 97.7%, and specificity of 99.1%. To highlight the important regions in the X-ray images which directs the model to its decision/prediction, we adopted the Gradient Class Activation Map (Grad-CAM). The generated heat maps from the Grad-CAM were then compared with the annotated X-ray images by board-certified radiologists. Results showed that the findings strongly correlate with clinical evidence. For practical deployment, we implemented the trained model in edge devices (NCS2) and this has achieved an improvement of 90% in inference speed compared to CPU. This shows that the developed model has the potential to be implemented on the edge, for example in primary care clinics and rural areas which are not well-equipped or do not have access to stable internet connections.

Forecasts of commodity prices are vital issues to market participants and policy-makers. Those of cooking section oil are of no exception, considering its importance as one of main food resources. In the present study, we assess the forecast problem using weekly wholesale price indices of canola and soybean oil in China during January 1, 2010–January 3, 2020, by employing the non-linear auto-regressive neural network as the forecast tool. We evaluate forecast performance of different model settings over algorithms, delays, hidden neurons, and data splitting ratios in arriving at the final models for the two commodities, which are relatively simple and lead to accurate and stable results. Particularly, the model for the price index of canola oil generates relative root mean square errors of 2.66, 1.46, and 2.17% for training, validation, and testing, respectively, and the model for the price index of soybean oil generates relative root mean square errors of 2.33, 1.96, and 1.98% for training, validation, and testing, respectively. Through the analysis, we show usefulness of the neural network technique for commodity price forecasts. Our results might serve as technical forecasts on a standalone basis or be combined with other fundamental forecasts for perspectives of price trends and corresponding policy analysis.

Community search enables personalized community discovery and has wide applications in real-life scenarios. Existing attributed community search algorithms use personalized information provided by attributes to locate desired community. Though achieved promising results, existing works suffer from two major limitations: (i) the precision of the algorithm decreases significantly when the seed comes from the boundary regions of the community. (ii) Most attributed community search methods mainly take the attribute information as edge weights to reveal semantic strength (e.g., attribute similarity, attribute distance, etc.), but largely ignore that attribute may serve as heterogeneous vertex. To make up for these deficiencies, in this paper, we propose a novel two-stage attributed community search method with seed replacement and joint random walk (SRRW). Specifically, in the seed replacement stage, we replace the initial query node with a core node; in the random walk stage, attributes are taken as heterogeneous nodes and the augmented graph is modeled based on the affiliation of the attributes via an overlapping clustering algorithm. And finally, a joint random walk is performed on the augmented graph to explore the desired local community. We conduct extensive experiments on both synthetic and real-world benchmarks, demonstrating its effectiveness for attributed community search.

The paper analyzes observations using a logic-based numerical methodology in Python. The Logical Analysis of Data (LAD) specializes in selecting a minimal number of features and finding unique patterns within it to distinguish ‘positive’ from ‘negative’ observations. The Python implementation of the classification model is further improved by introducing adaptations to pattern generation techniques. Finally, a case study of the Power Attack Systems Dataset used to improvise Smart Grid technology is performed to explore real-life applications of the classification model and analyze its performance against commonly used techniques.

In machine learning, a classifier has a certain learning curve i.e. the curve of the error/success probability as a function of the training set size. Finding the learning curve for a large interval of sizes takes a lot of processing time. A better method is to estimate the error probabilities only for few minimal sizes and use the pairs size-estimate as data points to model the learning curve. Searchers have tested different models. These models have certain parameters and are conceived from curves that only have the general aspect of a real learning curve. In this paper, we propose two new models that have more parameters and are conceived from real learning curves of nearest neighbour classifiers. These two main differences increase the chance for these new models to fit better the learning curve. We test these new models on one-input and two-class nearest neighbour classifiers.

The resolvent operator and the corresponding Green’s function occupy a central position in the realms of differential and integral equations, operator theory, and in particular the modern physics. However, in the field of machine learning, when confronted with the complex and highly challenging learning tasks from the real world, the prowess of Green’s function of resolvent is rarely explored and exploited. This paper aims at innovating the conventional translation-invariant kernels and rotation-invariant kernels, through theoretical investigation into a new view of constructing kernel functions by means of the resolvent operator and its Green’s function. From the practical perspective, the newly developed kernel functions are applied for robust signal recovery from noise corrupted data in the scenario of linear programming support vector learning. In particular, the monotonic and non-monotonic activation functions are used for kernel design to improve the representation capability. In this manner, a new dimension is given for kernel-based robust sparse learning from the following two aspects: firstly, a new theoretical framework by bridging the gap between the mathematical subtleties of resolvent operator and Green’s function theory and kernel construction; secondly, a concretization for the fusion between activation functions design in neural networks and nonlinear kernels design. Finally, the experimental study demonstrates the potential and superiority of the newly developed kernel functions in robust signal recovery and multiscale sparse modeling, as one step towards removing the apparent boundaries between the realms of modern signal processing and computational intelligence.

This paper addresses the issues concerning the rescheduling of a static timetable in case of a disaster, encountered in a large and complex railway network system. The proposed approach tries to modify the existing schedule to minimise the overall delay of trains. This is achieved by representing the rescheduling problem in the form of a Petri-Net and the highly uncertain disaster recovery time in such a model is handled as Markov decision processes (MDP). For solving the rescheduling problem, a distributed constraint optimisation (DCOP)-based strategy involving the use of autonomous agents is used to generate the desired schedule. The proposed approach is evaluated on the real-time data set taken from the Eastern Railways, India by constructing various disaster scenarios using the Java Agent DEvelopment Framework (JADE). The proposed framework, when compared to the existing approaches, substantially reduces the delay of trains after rescheduling.

Due to their dependence on a task-specific reward function, reinforcement learning agents are ineffective at responding to a dynamic goal or environment. This paper seeks to overcome this limitation of traditional reinforcement learning through a task-agnostic, self-organising autonomous agent framework. The proposed algorithm is a hybrid of TMGWR for self-adaptive learning of sensorimotor maps and value iteration for goal-directed planning. TMGWR has been previously demonstrated to overcome the problems associated with competing sensorimotor techniques such SOM, GNG, and GWR; these problems include: difficulty in setting a suitable number of neurons for a task, inflexibility, the inability to cope with non-markovian environments, challenges with noise, and inappropriate representation of sensory observations and actions together. However, the binary sensorimotor-link implementation in the original TMGWR enables catastrophic forgetting when the agent experiences changes in the task and it is therefore not suitable for self-adaptive learning. A new sensorimotor-link update rule is presented in this paper to enable the adaptation of the sensorimotor map to new experiences. This paper has demonstrated that the TMGWR-based algorithm has better sample efficiency than model-free reinforcement learning and better self-adaptivity than both the model-free and the traditional model-based reinforcement learning algorithms. Moreover, the algorithm has been demonstrated to give the lowest overall computational cost when compared to traditional reinforcement learning algorithms.

Machine learning approaches can serve as powerful tools in the machining optimization process. Criteria, such as accuracy and stability, are important to consider when choosing among different models. For the industrial application, it also is essential to balance cost, applicabilities, and ease of implementations. Here, we develop Gaussian process regression models for predicting the main cutting force (R) and its components in three directions of the coordinate system ((F_{x}), (F_{y}), and (F_{z})) based on two predictors: the depth of cut ((a_{p})) and the feed rate (f) in milling processes of functionally graded materials. The model performance shows high accuracy and stability, and the models are thus promising for estimating the cutting force and its component in a fast, cost effective, and robust fashion.

It is an established fact that the genuineness of facial micro-expression is an effective means for estimating concealed emotions (Li et al. in Micro-expression recognition under low-resolution cases. SciTePress, Science and Technology Publications, Setúbal, 2019). Conventionally, analysis of these expressions has been performed using high resolution images which are ideal cases. However, in a real-world scenario, capturing expressions with high resolution images may not always be possible particularly using low-cost surveillance cameras. Faces captured using such cameras are often very tiny and of poor resolution. Due to the loss of discriminative features these images may not be of much use particularly for identifying certain minute facial details. To make these images useful, enhancing the textural information becomes essential and super-resolution algorithms can be ideal to achieve this. In this work, we utilize algorithms based on deep learning and generative adversarial network for transforming low-resolution micro-expression images into super-resolution images and examine their fitness particularly for micro-expression recognition. The proposed approach is tested on simulated dataset obtained from two popular spontaneous micro-expression datasets namely CASME II and SMIC-VIS; the experimental results demonstrate that the method achieved favourable results with the best recognition performance recorded as 61.63%. The significance of this work is: first, it thoroughly investigates reconstruction performance of several deep learning super-resolution algorithms on simulated low-quality micro-expression images; second, it provides a comprehensive analysis of the results obtained employing these reconstructed images to determine their contribution in addressing image quality issues specifically for micro-expression recognition.