Search

With the wide application of the shield tunneling method in tunnel engineering, the untimely and incorrect attitude control of shield systems has become an essential factor affecting the quality of shield tunneling. The use of multi-agent deep reinforcement learning (MADRL) algorithm to learn correct and timely shield attitude correction policies is beneficial for handling shield attitude control tasks which are complex and separable. However, the sensing data of the shield systems are complicated and widely distributed, which affects the learning efficiency of the agents. To resolve this problem, this paper proposes a MADRL framework based on state classification and assignment (SCA) for adaptive shield attitude control. SCA-MADRL uses a clustering algorithm to classify states and uses a planning model to make the agent's learning space a consistent state space so that each agent can learn efficiently in its own learning space. The state classification algorithm in the framework can improve the learning efficiency of the agent, and the proposed assignment algorithm can further enhance the learning effect when reusing the agent. This paper verifies the necessity and effectiveness of using task decomposition to implement MADRL for equipment control and provides a practical framework for automated and intelligent shield systems. [ABSTRACT FROM AUTHOR]

With the advancement of the energy transition, renewable power-dominated new power systems are faced with unprecedented technical challenges. Faced with the increasing uncertainty brought by renewables and the exponential growth of the number of flexible resources, data-driven methods have shown great potential in recent years due to their adaptability in scenarios with massive decisions and high efficiency. Therefore, this paper proposes a data-driven look-ahead economic dispatch model with full consideration of N-1 outage contingency based on reinforcement learning and a deep deterministic policy gradient (DDPG) algorithm. We first use the domain knowledge of dynamic economic dispatch to formulate the reinforcement learning model, and then utilize an improved DDPG algorithm with mode classification for the training and validation. The N-1 outage contingency is considered based on the idea of transfer learning to reduce the size of N-1 scenarios and further improve the efficiency. Case studies based on the IEEE30-bus and SG126-bus systems validate the effectiveness of the proposed model and algorithm, with superior performance in terms of security and economy compared to the conventional model-based method. • This manuscript focuses on the look-ahead economic dispatch problem considering N-1 contingencies based on the idea of reinforcement learning and transfer learning. The major contributions are summarized as follows: • An efficient parallel computing framework is established to fine-tune the dispatch schemes over the upcoming hours. • A new single-interval training scheme is designed to improve the efficiency of the proposed method in large systems. • An improved DDPG algorithm is proposed with an expert control strategy to improve the reliability of neural network training. [ABSTRACT FROM AUTHOR]

Artificial intelligence is steadily increasing its impact on everyday life. Therefore, the societal issues of artificial intelligence have become an important concern in the AI research. The presence of data that reflects human biases towards historically discriminated groups defined by sensitive features such as race and gender, results in machine learning models which discriminate against these groups. In order to tackle the impact of bias in data, researchers developed a variety of specialized machine learning algorithms which are able to satisfy different fairness constraints imposed on the model. Group fairness constraints do not fit standard machine learning formulations easily due to their non-differentiable nature. In this paper we developed a technique for learning a fair classifier by Monte Carlo policy gradient method which naturally deals with such non-differentiable constraints. Our methodology focuses on direct optimization of both group fairness metric and predictive performance of the model. In addition, we propose two different variance reduction techniques of gradient estimation. We compare our models to seven other related and state-of-the-art models and demonstrate that they are able to achieve better trade-off between accuracy and unfairness. To the best of our knowledge, this is the first fair classification algorithm which solves the issue of non-differentiable constraints by reinforcement learning techniques. [ABSTRACT FROM AUTHOR]