جستجوی مقالات مرتبط با کلیدواژه « Deep reinforcement learning » در نشریات گروه « فنی و مهندسی »

Web applications (apps) are integral to our daily lives. Before users can use web apps, testing must be conducted to ensure their reliability. There are various approaches for testing web apps. However, they still require improvement. In fact, they struggle to achieve high coverage of web app functionalities. On the one hand, web apps typically have an extensive state space, which makes testing all states inefficient and time-consuming. On the other hand, specific sequences of actions are required to access certain functionalities. Therefore, the optimal testing strategy extremely depends on the app’s features. Reinforcement Learning (RL) is a machine learning technique that learns the optimal strategy to solve a task through trial-and-error rather than explicit supervision, guided by positive or negative reward. Deep RL extends RL, and exploits the learning capabilities of neural networks. These features make Deep RL suitable for testing complex state spaces, such as those found in web apps. However, modern approaches support fundamental RL. We have proposed WeDeep, a Deep RL testing approach for web apps. We evaluated our method using seven open-source web apps. Results from experiments prove it has higher code coverage and fault detection than other existing methods

Composable life under the extensive global warming of the Earth encourages the progress of renewable energy devices and the adoption of new technologies, such as artificial intelligence. Regarding enormous potential of wave energy and its consistency, wave energy converter (WEC) plays vital role in uniform energy harvesting field. In this paper, the significant environmental changes in the ocean prompt us to propose an intelligent feedback control system to mitigate the impact of disturbances and variable wind effects on the efficacy of WECs. Deep reinforcement learning (DRL), as a powerful machine intelligence technique, is capable of identifying WECs as black-box models. Therefore, based on the DRL model, the disturbance and unmeasured state variables are simultaneously estimated in the extended state observer section. Leakage in identification data and real-time application requirements of limited number of layers in the deep neural networks are compensated by implementation of immersion and invariance-based extended state observer which improves coping with the unwanted exogenous noises as well. In the overall intelligent control system, the estimated parameters are inputted into the DRL as the actor-critic networks. The initial actor network is responsible for predicting the control action, while the subsequent critic network determines the decision criterion for evaluating the accuracy of the actor's estimated amount. Next, the output value of the critic stage is backpropagated through the layers to update the network weights. The simulation test results in MATLAB indicate the convergence of unmeasured parameters/states to the corresponding true values and the significance of newly designed intelligent DRL method.

 Web application (app) exploration is a crucial part of various analysis and testing techniques. However, the current methods are not able to properly explore the state space of web apps. As a result, techniques must be developed to guide the exploration in order to get acceptable functionality coverage for web apps. Reinforcement Learning (RL) is a machine learning method in which the best way to do a task is learned through trial and error, with the help of positive or negative rewards, instead of direct supervision. Deep RL is a recent expansion of RL that makes use of neural networks’ learning capabilities. This feature makes Deep RL suitable for exploring the complex state space of web apps. However, current methods provide fundamental RL. In this research, we offer DeepEx, a Deep RL-based exploration strategy for systematically exploring web apps. Empirically evaluated on seven open-source web apps, DeepEx demonstrated a 17% improvement in code coverage and a 16% enhancement in navigational diversity over the stateof-the-art RL-based method. Additionally, it showed a 19% increase in structural diversity. These results confirm the superiority of Deep RL over traditional RL methods in web app exploration.

A closed-loop control framework is developed for the co-flow jet (CFJ) airfoil by combining the numerical flow field environment of a CFJ0012 airfoil with a deep reinforcement learning (DRL) module called tensorforce integrated in Python. The DRL agent, which is trained through interacting with the numerical flow field environment, is capable of acquiring a policy that instructs the mass flow rate of the CFJ to make the stalled airfoil at an angle of attack (AoA) of 18 degrees reach a specific high lift coefficient set to 2.0, thereby effectively suppressing flow separation on the upper surface of the airfoil. The subsequent test shows that the policy can be implemented to find a precise jet momentum coefficient of 0.049 to make the lift coefficient of the CFJ0012 airfoil reach 2.01 with a negligible error of 0.5%. Moreover, to evaluate the generalization ability of the policy trained at an AoA of 18 degrees, two additional tests are conducted at AoAs of 16 and 20 degrees. The results show that, although using the policy gained under another AoA cannot help the lift coefficient of the airfoil reach a set target of 2 accurately, the errors are acceptable with less than 5.5%, which means the policy trained under an AoA of 18 degrees can also be applied to other AoAs to some extent. This work is helpful for the practical application of CFJ technology, as the closed-loop control framework ensures good aerodynamic performance of the CFJ airfoil, even in complex and changeable flight conditions.

Contemporary machine learning models, like deep neural networks, require substantial labeled datasets for proper training. However, in areas such as natural language processing, a shortage of labeled data can lead to overfitting. To address this challenge, data augmentation, which involves transforming data points to maintain class labels and provide additional valuable information, has become an effective strategy. In this paper, a deep reinforcement learning-based text augmentation method for sentiment analysis was introduced, combining reinforcement learning with deep learning. The technique uses Deep Q-Network (DQN) as the reinforcement learning method to search for an efficient augmentation strategy, employing four text augmentation transformations: random deletion, synonym replacement, random swapping, and random insertion. Additionally, various deep learning networks, including CNN, Bi-LSTM, Transformer, BERT, and XLNet, were evaluated for the training phase. Experimental findings show that the proposed technique can achieve an accuracy of 65.1% with only 20% of the dataset and 69.3% with 40% of the dataset. Furthermore, with just 10% of the dataset, the method yields an F1-score of 62.1%, rising to 69.1% with 40% of the dataset, outperforming previous approaches. Evaluation on the SemEval dataset demonstrates that reinforcement learning can efficiently augment text datasets for improved sentiment analysis results.

The hybrid electric train which operates without overhead wires or traditional power sources relies on hydrogen fuel cells and batteries for power. These fuel cell-based hybrid electric trains (FCHETs) are more efficient than those powered by diesel or electricity because they do not produce any tailpipe emissions making them an eco-friendly mode of transport. The target of this paper is to propose low-budget FCHETs that prioritize energy efficiency to reduce operating costs and minimize their impact on the environment. To this end, an energy management strategy [EMS] has been developed that optimizes the distribution of energy to reduce the amount of hydrogen required to power the train. The EMS achieves this by balancing battery charging and discharging. To enhance the performance of the EMS, proposes to use of a deep reinforcement learning (DRL) algorithm specifically the deep deterministic policy gradient (DDPG) combined with transfer learning (TL) which can improve the system's efficiency when driving cycles are changed. </strong>DRL-based strategies are commonly used in energy management and they suffer from unstable convergence, slow learning speed, and insufficient constraint capability. To address these limitations, an action masking technique to stop the DDPG-based approach from producing incorrect actions that go against the system's physical limits and prevent them from being generated is proposed. </strong> The DDPG+TL agent consumes up to 3.9% less energy than conventional rule-based EMS while maintaining the battery's charge level within a predetermined range. The results show that DDPG+TL can sustain battery charge at minimal hydrogen consumption with minimal training time for the agent.

The quality of the extracted features from a long-term sequence of raw prices of the instruments greatly affects the performance of the trading rules learned by machine learning models. Employing a neural encoder-decoder structure to extract informative features from complex input time-series has proved very effective in other popular tasks like neural machine translation and video captioning. In this paper, a novel end-to-end model based on the neural encoder-decoder framework combined with deep reinforcement learning is proposed to learn single instrument trading strategies from a long sequence of raw prices of the instrument. In addition, the effects of different structures for the encoder and various forms of the input sequences on the performance of the learned strategies are investigated. Experimental results showed that the proposed model outperforms other state-of-the-art models in highly dynamic environments.

Condition-based maintenance (CBM) involves making decisions on maintenance based on the actual deterioration conditions of the components. It consists of a chain of states representing various stages of deterioration and a set of maintenance actions. Therefore, condition-based maintenance is a sequential decision-making problem. Reinforcement Learning(RL) is a subfield of Machine Learning proposed for automated decision-making. This article provides an overview of reinforcement learning and deep reinforcement learning methods that have been used so far in condition-based maintenance optimization.

In recent years, exponential growth of communication devices in Internet of Things (IoT) has become an emerging technology which facilitates heterogeneous devices to connect with each other in heterogeneous networks. This communication requires different level of Quality-of-Service (QoS) and policies depending on the device type and location. To provide a specific level of QoS, we can utilize emerging new technological concepts in IoT infrastructure, software-defined network (SDN) and, machine learning algorithms. We use deep reinforcement learning in the process of resource management and allocation in control plane. We present an algorithm that aims to optimize resource allocation. Simulation results show that the proposed algorithm improved network performances in terms of QoS parameters, including delay and throughput compared to Random and Round Robin methods. Compared to similar methods the performance of the proposed method is also as good as the fuzzy and predictive methods.