Power distribution method in downlink NOMA of depth deterministic strategy gradient

Abstract

The invention discloses a power distribution method in a downlink NOMA system of a depth deterministic strategy gradient algorithm, and the method employs a double neural network structure and an experience pool playback mechanism, can effectively solve a problem related to a large-scale state action space, reduces the correlation between training samples, employs a deterministic strategy to select an action, and an action may be selected in a continuous action space. According to the algorithm, state information is used as input of a neural network, a state space, an action space and a rewardfunction are correspondingly designed according to a simulation downlink NOMA system situation, and signal to interference plus noise ratio information and rate information at the last moment are used as components of the state information at the current moment. According to the invention, the intelligent agent can learn and utilize the learned information to improve the behavior strategy more effectively, and an optimal power distribution strategy is obtained after multiple iterations. The method can effectively solve the problem of power distribution of multiple users in the downlink NOMA system, has good generalization performance under different user numbers and the transmitting power level of the base station, can effectively improve the rationality of power distribution, consumes less operation time, and effectively improves the efficiency of power distribution.

Description

technical field [0001] The invention relates to the field of NOMA resource allocation, in particular to a power allocation method in a downlink NOMA system based on a deep deterministic strategy gradient algorithm. Background technique [0002] With the continuous access of mobile terminal equipment and the continuous increase of user density in the wireless communication system, the amount of data in the communication system has shown an exponential growth, and the orthogonal multiple access technology has been unable to meet the needs of high system capacity. The fifth-generation mobile communication system emerged as the times require. The main focus of 5G technology is the improvement of data rate and the reduction of end-to-end delay, so as to adapt to the exponential growth of wireless business data volume. Non-orthogonal multiple access (NOMA) is considered to be a promising technology in the 5G communication system, which allows multiple users to communicate on the s...

AI Technical Summary

Problems solved by technology

[0004] Power allocation problems are mostly NP-hard (NP-hard), and have non-convexity, and it is not easy to find the optimal solution. Therefore, there are many research methods that use explicit or implicit optimization techniques, through iterative Calculation to find the optimal solution, the traditional method can effectively improve the system performance, but the numerical simulation method does not have an accurate system model, and multiple iterative calculations involve a lot of calculations and consume a lot of time, which cannot meet the requirements of the communication system. Real-time data processing requirements, and based on the traditional method does not have the ability of self-learning, unable to adapt to changeable and complex communication systems, so the processing of actual communication scenarios is not perfect

[0005] Following the traditional optimization algorithm, some scholars have proposed the use of deep learning technology to solve the power allocation problem in the NOMA system. Such methods use deep neural networks or other variants, and use supervised learning. Through multi-layer The neural network is used to extract data features and learn the mapping from data to labels. Compared with the multiple iterative calculations of the traditional power allocation method, the operation of the neural network is more efficient and less complex, but the training of the neural network requires a lot of preparation. Good sample data, but it is difficult to obtain perfect sample data in the communication system, and the supervised learning method requires a benchmark algorithm for training, and its performance will be limited by the benchmark algorithm

[0006] Compared with the supervised learning method in deep learning, reinforcement learning adopts a self-learning strategy. The agent continuously learns the observed environmental information, and constantly updates its own behavior selection strategy, and finally learns the optimal one. The optimal behavior control strategy, the Q-learning algorithm is the most classic reinforcement learning algorithm, but the traditional Q-learning algorithm suffers from the "curse of dimensionality", that is, it cannot deal with high-dimensional state-action space problems, and can only deal with discrete actions Space, powerless for continuous action space, deep reinforcement learning effectively solves the curse of dimensionality problem by using deep neural network instead of traditional Q-value function. The most widely used algorithm is Deep Q Network (DQN), but it also has problems that cannot be solved. Continuous action space and other issues, there are certain limitations

Images