In this article, we propose a heterogeneous federated learning for classification (HFLC) model, which divides features into sensitive and nonsensitive for the privacy concerns. The local-network is first trained on all features to achieve specificity, and the global-network is then trained on nonsensitive features. They are are the carriers of heterogeneous federated learning (HFL) to achieve generalization. The designed HFL is then conducted on global-networks to deal with the non-IID distributions of features and labels. Experimental results show that the proposed HFLC systems reduce the classification error from 29.5% to 36.7% compared to the traditional federated average; from 17.7% to 21.4% compared to existing heterogeneous federated algorithms, and from 2.2% to 5.0% compared to state-of-the-art systems.

Detection and prediction of failures in Automated Guided Vehicles (AGV) are essential for the uninterrupted operation of production plants. Anomaly detection is usually achieved by comparing expected measurement values with actual observations. Thus, it is crucial to predict telemetry signals properly. In this paper, we research the prediction of energy consumption using state-of-the-art Artificial Neural Networks architectures (SCINet) compared with other Recurrent Neural Network (RNN) approaches on the data streams acquired from CoBotAGV. We especially focus on the possibility of applying feature weighting. We show that it can improve prediction capabilities. We also investigate resource utilization in terms of time to fit the embedded AGV environment.

Smart cities require the use of many different types of sensors to make the communication, and distance sensors are one of the most commonly used elements in transportation systems and related infrastructures. The introduction of increasingly advanced autonomous systems in many areas of smart cities requires high measurement precision of the sensors used. High precision is essential for proper operation, long-term use, and safety in machine-to-machine or machine-to-human interactions. This paper presents a comparison of the accuracy of distance measurements for two commercially available single-beam LiDARs and two ultrasonic sensors. The aim of the research was to develop a calibration method in order to improve the accuracy of distance sensors. Based on the collected distance measurements, the sensors were calibrated using selected machine learning algorithms. The results of the experiments show the effectiveness of the proposed calibration methods, which yield an average mean absolute error (MAE) of 1.76 E-05 meters (m) and a root mean square error (RMSE) of 1.33 E-04 m for the tested sensors.

Medical internet of things leads to revolutionary improvements in medical services, also known as smart healthcare. With the big healthcare data, data mining and machine learning can assist wellness management and intelligent diagnosis, and achieve the P4-medicine. However, healthcare data has high sparsity and heterogeneity. In this paper, we propose a Heterogeneous Transferring Prediction System (HTPS). Feature engineering mechanism transforms the dataset into sparse and dense feature matrices, and autoencoders in the embedding networks not only embed features but also transfer knowledge from heterogeneous datasets. Experimental results show that the proposed HTPS outperforms the benchmark systems on various prediction tasks and datasets, and ablation studies present the effectiveness of each designed mechanism. Experimental results demonstrate the negative impact of heterogeneous data on benchmark systems and the high transferability of the proposed HTPS.

Renewable energy generation has attracted the interest of researchers, but it is volatile, and management systems are vulnerable to malicious attacks. Therefore, security issues are of paramount importance for energy management systems. In this paper, we propose a secure Q-learning- based energy network management system (SQEMS), which consists of an anomaly detection module, a fuzzy control module to mitigate attacks, and a decision-making module to manage the energy grid. Experimental results show that the proposed anomaly detection module has excellent performance on malicious suppliers attacks (MS), and the fuzzy control module can further mitigate the negative effects of false predictions. The robustness analysis shows the effectiveness, robustness, and transferability in anomaly detection and energy management.

With the rapid development of quantitative trading, stock selection is an ongoing task that requires consideration of the characteristics of stocks and investment strategies. Fuzzy set theory is excellent for modeling and describing abstract characteristics to establish the link between stocks and strategies. In this article, we present our idea for a stock selection system called TripleS. It is based on the suitability index (SI) derived from fuzzy-set theory. The TripleS relies on the position size to extract the stock characteristics SI that describes not only the characteristics of each stock but also the extent to which they are suitable for certain strategies. Then, the SI undergoes a transformation process in the hands of the developed fuzzy-set modules for stock selection and investment. There are some other proposed methods for stock selection respectively called TripleS-T1, TripleS-T2, TripleS-T2U, TripleS-EF, and TripleS-EFT. Experimental results show that the proposed TripleS has minor improvement in precision but significant improvement in investment performance. The TripleS-EFT has outstanding profitability, the highest annual return in all strategies, and the highest Sharpe ratio in two of the four strategies, and achieves an annual return of 10.01% and a Sharpe ratio of 1.032, which significantly outperform the benchmark system. The TripleS-EFT also significantly outperforms the state-of-the-art fuzzy-based system in measuring annual return and even has double the annual return in several strategies.

The use of Artificial Intelligence (AI) to support the Automated Guided Vehicles (AGV) that are used by industry poses a number of challenges that are specific to smart internal logistics systems that are necessary for agile manufacturing. On the one hand, it might seem that experience with the autonomous navigation system that are used in autonomous vehicles can be easily transferred to AGV. However, in this paper, the authors highlight specific problems that are associated with the navigation system of AGV, which has to reflect its operation in an industrial environment with high level of interaction with other production systems and human staff. On the other hand, it may seem that the wealth of experience from using AI in smart manufacturing can be easily transferred to the use of AGV. However, the authors show that although AGV are production tools, the challenges that are associated with the use of AI can significantly differ from other smart manufacturing areas. The number of challenges that are specific to use of AI for AGV is also discussed. This paper systematizes these challenges and discusses the most promising AI methods that can be used for the internal logistics systems that are based on AGV.

In this paper, we propose a Secure Energy Management System (SEMS) with anomaly detection and Q-Learning decision modules for Automated Guided Vehicles (AGV). The anomaly detection module is a multi-task learning network to simultaneously classify suppliers and predict the real supply quantities. The Q-learning decision module can then determine operating reserve and subsidies to manage the energy grid. Experimental results illustrate that the proposed anomaly detection module has an excellent performance in classifying malicious suppliers, excels at shaping supply distribution, and outperforms the existing benchmark systems.

In this research, we present a Q-learning based energy management system (DEQEMS) that is able to make decisions by using unique states and intuitive actions while maintaining a high degree of interpretability. The results of the experiments show that the DEQEMS reduces the number of days required for convergence to 633, with a mean absolute error (MAE) of supply distribution of 6.7%. This is a 63% and 71% reduction, respectively, compared to the conventional system, and a 34% and 21% reduction, respectively, compared to a state-of-the-art system. The experimental results demonstrate not only the usefulness and feasibility of the DEQEMS, but also its resilience with outstanding and consistent performance under a wide range of conditions.

With the rapid development of quantitative trading, stock selection is an enduring task, that requires consideration of the characteristic of stocks and investment strategies. Fuzzy-set theory is outstanding in modeling abstract characteristics, which is suitable for building the connection between stocks and strategies. In this paper, we propose a stock selection system (SSS) based on the suitability index (SI) and fuzzy-set theory. SSS utilizes position sizing to extract stock characteristics, SI, which expresses not only the characteristics of stocks but also the suitability for strategies. The SI is then transformed by the designed fuzzy-set modules for stock selection and investment. Several stock selection mechanisms are proposed, denoted as SSS-T1, SSS-T2, SSS-WT1, and SSS-WT2. Experimental results show that SSS has a slight improvement in precision, but leads to a significant improvement in investment. The proposed SSS-WT1 and SSS-WT2 systems achieve the highest annual returns of 6.2% and the highest Sharpe ratios of 0.924, outperforming the benchmark with the annual return of 1.9% and the Sharpe ratio of 0.201.

Trading strategies can be divided into two categories, i.e., those with momentum characteristic and those that appear contrarian. The characteristics of trading strategies have been widely studied; however, there has been relatively little work on the characteristics of stocks. Furthermore, there is no standard approach to the classification of stocks in terms of momentum and contrarian. This article presents a fuzzy momentum contrarian uncertain characteristic system for the classification and quantification of stock characteristics. Random trading, stop-loss, and take-profit mechanisms are first used to identify characteristics, and then, a novel profitability index with a type-2 fuzzy set module is used to quantify them. In the experiments, 41 stocks on the Taiwan 50 index were deemed suitable for momentum strategies, whereas nine stocks were deemed suitable for contrarian strategies. An uphill relationship between profitability index and trading performance is observed, which produced correlation coefficients of 0.148–0.539 and a classification accuracy of 52.0–60.0%. However, the proposed system greatly improved classification performance, resulting in correlation coefficients of 0.572–0.722 with an accuracy of 63.6–84.5%. In the real-world application, the proposed system outperforms the benchmark among all datasets and increases the profitability by 1.5 times on the Taiwan 50 dataset. These results clearly demonstrate the efficiency of the proposed system in the quantification and classification of stocks suited to momentum- and contrarian-type trading strategies and also in the real-world applications.

Abstract Portfolio management involves position sizing and resource allocation. Traditional and generic portfolio strategies require forecasting of future stock prices as model inputs, which is not a trivial task since those values are difficult to obtain in the real-world applications. To overcome the above limitations and provide a better solution for portfolio management, we developed a Portfolio Management System (PMS) using reinforcement learning with two neural networks (CNN and RNN). A novel reward function involving Sharpe ratios is also proposed to evaluate the performance of the developed systems. Experimental results indicate that the PMS with the Sharpe ratio reward function exhibits outstanding performance, increasing return by 39.0% and decreasing drawdown by 13.7% on average compared to the reward function of trading return. In addition, the proposed model is more suitable for the construction of a reinforcement learning portfolio, but has 1.98 times more drawdown risk than the . Among the conducted datasets, the PMS outperforms the benchmark strategies in TW50 and traditional stocks, but is inferior to a benchmark strategy in the financial dataset. The PMS is profitable, effective, and offers lower investment risk among almost all datasets. The novel reward function involving the Sharpe ratio enhances performance, and well supports resource-allocation for empirical stock trading.

Environmental protection is an important issue in recent decades, and renewable energy is an ideal solution for eco-friendly power generation. Solar-power generation is a popular renewable energy with low cost and small environmental footprint, which leads to exponential growth and high industrial investment. A mature solar business model has been established, but some uncertainties hinder the development, especially when focusing on the lack of solar-radiation. To address these issues, in this article we propose a hedging system to hedge the low-radiation risk for solar-investors through the designed IoT-based data, edge-based models for predicting solar-radiation as well as hedging options. Our experimental results show that the edge-based predictive models can obtain an R-squared value of 0.841 and a correlation coefficient of 0.917. For binary options designed in the hedging system, the broker can obtain stable payoffs with the highest Sharpe ratio of 3.354, and the investors can obtain large payoffs during low-radiation. Our simulation results show the effectiveness of the proposed hedging system for investors (buyer-side), simultaneously, present the motivation of the broker (seller-side) to join the designed hedging system utilized in solar-power generation.

Opening range breakout (ORB) is a well-known intraday trading strategy via technical analysis. ORB lacks robustness against market uncertainties (e.g., information from contradictory sources), and does not consider all relevant market characteristics. Furthermore, the closing strategies in generic ORB are not well defined. In this study, we developed an evolutionary ORB-based model, which utilized historical data to optimize thresholds in order to enhance profitability, and developed protective closing strategies aimed at to prevent unacceptable losses. Selecting appropriate thresholds and parameters for ORB is a non-trivial task, due to the fact that the search space exceeds sixty-five thousand options. We used evolutionary computation to derive rational strategies and parameters for ORB. The proposed framework based on a genetic algorithm optimizes the parameters related to threshold selection and protective closing strategies. In experiments, this resulted in annual returns of 9.3% (representing a 2.8% improvement over the original strategy) and Sharpe ratio of 2.5 (an improvement of 1.0), while reducing the maximum drawdown by half. The proposed scheme also reduced computational overhead by 89% compared to a grid search.

Opening range breakout (ORB) is a well-known intraday trading strategy that generates trading signals through technical analysis; however, ORB does not make full use of market characteristics and does not define closing strategy. These problems make the ORB strategy not stable or robust enough. In this paper, we adjust thresholds through historical data to enhance profitability, and design protective closing strategy to prevent unacceptable losses. However, there are numerous parameters combinations, and the solution space is approximately 214. Therefore, we implement genetic algorithm to improve the efficiency and rationality of parameter selection. We found that the performance of the GAORB_Sharpe with stop-loss mechanism is outstanding. The strategy we proposed can generate 9.303% annual return and 15.716% Sharpe ratio, which is 2.5% and 5% more than the original strategy, and cut the maximum drawdown by half. Further, it can save 90% of the computation by genetic algorithm. In summary, we recommend adjusting the threshold and implementing stop-loss mechanism to the ORB strategy, and selecting parameters through genetic algorithms to improve overall performance.

Opening Range Breakout (ORB) is a renowned technical analysis strategy in which two pre-determined thresholds are set up in the early stage after market opening to determine the direction of investment for the day. In the literature, ORB has been shown to produce significant profits on several stock markets. In this paper, we present novel ORB algorithms based on deep learning which take into account historical trends and price movement during the early market interval. The proposed scheme uses a multi-label classification framework with multi-layer perceptron neural models and convolutional neural network models to predict the most profitable thresholds. To generate ground truth values for evaluation, we compared two labeling methods (one-hot encoding and distributed encoding). In experiments based on an empirical dataset, the proposed strategy earned profits of 8126, annual returns of 14.003%, and a Sharpe ratio of 1.376. The proposed scheme outperformed the original ORB strategy by nearly 2 times on metrics mentioned above and decreased maximum drawdown by more than 60%. Experiment results revealed that the uppermost and lowermost classes of thresholds accounted for the majority of the predicted results. In other words, taking a long position at a lower boundary and a short position at a higher boundary increased the likelihood of generating higher profits, while reducing exposure to risk.

Portfolio management is a critical issue which should be skilled by position sizing and resource allocation. Traditional and generic portfolio strategies require to forecast the future stocks prices as the model inputs, which is not a trivial task in the real-world applications. To solve the above limitations and provide a better solution for the portfolio management to the inventors, we then develop a portfolio management system (PMS) with equity market neutral strategy in reinforcement learning. A novel reward function involving Sharpe ratio is also designed to evaluate the performance of the developed systems. Experimental results indicate that the PMS with Sharpe ratio reward function has the outstanding performance, and increase the return 39.0% and decrease the drawdown of 13.7% on average than that with reward function of trading return. In addition, the developed PMS_CNN model is more suitable and profitable to construct RL portfolio, but has a 1.98 times more drawdown risk than the PMS_RNN. Overall, the proposed PMS outperforms the benchmark strategies in the measurements of total return and Sharpe ratio. The PMS is profitable and effective with lower investment risk, and the novel reward function by involving Sharpe ratio really enhances the performance, and well support the resource-allocation in the empirical stock trading.

Portfolio management and money management have always been important issues for investors and researchers in the financial field. The Kelly criterion is a theoretical approach of money management, and is a mathematical method for optimizing long-term expected return. Kelly criterion requires the future outcomes distribution as input, which can be predicted through the techniques of machine learning (ML). With the revolutionary growth of the amount of information, big data is the key to boost ML prediction, therefore, we introduce a general Kelly framework, including the strength of Kelly, ML, and big data. In addition, we propose the Kelly stationary index (KSI) to quantify the stationarity of the stock's outcomes distribution, which will affect the trading period and forecasting frequency. We calculate the KSI of each constituent stock of Taiwan's 50, and apply the Kelly criterion strategy to verify the effectiveness of KSI. The experimental results show that there is a moderate downhill relationship between the strategy performance and KSI with the -0.591 of correlation coefficient. It also indicates that the closer the estimated distribution is to the actual distribution, the higher the expected profit. In the future, we will use KSI for money management, strategy development, and apply KSI into the general Kelly framework.

Opening Range Breakout (ORB) is a fairly intraday trading strategy. We set the resistance and the support levels by the price in opening interval to follow the trend in the futures market. However, such kind of strategies is not profitable for most commodities in recent years in the changing market. In this paper, we attempt to improve the original ORB strategy by considering the effect of trends continuity on the event. We adjust the predetermined threshold for upper bound and lower bound. This strategy is called Threshold Adjusting ORB or TA_ORB. We implement this modified ORB strategy on the Taiwan Index Futures from 2008 to 2012. Compared with the original ORB strategy, we got 145.98% return in 2008 (bear market), 81.86% return in 2009 (bull market) and 32.25% annual return in 2008-2012 (five-year period) which are 4.0 times, 1.4 times, and 2.6 times more than original ORB, respectively. TA_ORB performs outstanding in large fluctuation, especially in the bear market. Performance can verify that the observations of TA_ORB improve the stability of the breakthrough signal, enhance the return, and reduce strategic risk. Further, we plan to use neural network to make more precise predictions and implement these strategies in different commodities.

The state-of-the-art models for semantic image segmentation usually contain a convolutional neural network (CNN) and a conditional random field (CRF). As a predictor, existing CNN techniques can generate a dense prediction result but may generate obvious boundary errors at the same time. As a refinement model, CRF improves the CNN outcomes by forcing the consistency of local labels. However, the use of CRF may cause fragmentation effect around object boundaries. In this paper, we propose the use of a so-called iterative contraction and merging (ICM) process to facilitate the semantic segmentation process. Guided by the high-level information from CNN, the ICM process is used as a tool to grow image segments in a bottom-up way and to produce more accurate outcomes in an iterative way. The ICM process can faithfully preserve the boundary information and maintain the consistency of local labels. Our experimental results demonstrate that the performance of the proposed approach is comparable to the state-of-the-art models but with more accurate boundaries.

In this paper, we propose a new framework for hierarchical image segmentation based on iterative contraction and merging. In the proposed framework, we treat the hierarchical image segmentation problem as a sequel of optimization problems, with each optimization process being realized by a contraction-and-merging process to identify and merge the most similar data pairs at the current resolution. At the beginning, we perform pixel-based contraction and merging to quickly combine image pixels into initial region-elements with visually indistinguishable intra-region color difference. After that, we iteratively perform region-based contraction and merging to group adjacent regions into larger ones to progressively form a segmentation dendrogram for hierarchical segmentation. Comparing with the state-of-the-art techniques, the proposed algorithm can not only produce high-quality segmentation results in a more efficient way, but also keep a lot of boundary details in the segmentation results.