A Novel Q-stem Connected Architecture for Beyond-Diagonal Reconfigurable Intelligent Surfaces

Beyond-diagonal reconfigurable intelligent surface (BD-RIS) has garnered significant research interest recently due to its ability to generalize existing reconfigurable intelligent surface (RIS) architectures and provide enhanced performance through flexible inter-connection among RIS elements. However, current BD-RIS designs often face challenges related to high circuit complexity and computational complexity, and there is limited study on the trade-off between system performance and circuit complexity. To address these issues, in this work, we propose a novel BD-RIS architecture named Q-stem connected RIS that integrates the characteristics of existing single connected, tree connected, and fully connected BD-RIS, facilitating an effective trade-off between system performance and circuit complexity. Additionally, we propose two algorithms to design the RIS scattering matrix for a Q-stem connected RIS aided multi-user broadcast channels, namely, a low-complexity least squares (LS) algorithm and a suboptimal LS-based quasi-Newton algorithm. Simulations show that the proposed architecture is capable of attaining the sum channel gain achieved by fully connected RIS while reducing the circuit complexity. Moreover, the proposed LS-based quasi-Newton algorithm significantly outperforms the baselines, while the LS algorithm provides comparable performance with a substantial reduction in computational complexity.

Deep reinforcement learning for portfolio management based on the empirical study of chinese stock market

The objective of this paper is to verify that current cutting-edge artificial intelligence technology, deep reinforcement learning, can be applied to portfolio management. We improve on the existing Deep Reinforcement Learning Portfolio model and make many innovations. Unlike many previous studies on discrete trading signals in portfolio management, we make the agent to short in a continuous action space, design an arbitrage mechanism based on Arbitrage Pricing Theory,and redesign the activation function for acquiring action vectors, in addition, we redesign neural networks for reinforcement learning with reference to deep neural networks that process image data. In experiments, we use our model in several randomly selected portfolios which include CSI300 that represents the market's rate of return and the randomly selected constituents of CSI500. The experimental results show that no matter what stocks we select in our portfolios, we can almost get a higher return than the market itself. That is to say, we can defeat market by using deep reinforcement learning.