Research On Performance Optimization Of Rate Splitting Multiple Access-based Satellite Networks

As the number of wireless users and new services continuously grows,the greater demands for the 6th generation mobile communication systems(6G)are presented,in which extremely wide coverage requires the cooperation of satellites and terrestrial networks.However,the existing multiple access technologies will not meet the requirements of massive access from rapid growth of users and devices in the satellite networks.New multiple access technologies in satellite communications and their optimization methods become a research hotspot.This dissertation proposes several optimization algorithms with stronger learning ability comparing to conventional ones,which are more suitable for the next generation mobile communication systems.The main innovations are summarized below.1.An improved deep unfolding algorithm is proposedThis dissertation proposes an improved deep unfolded weighted minimum mean square error(WMMSE)algorithm to optimize the sum rate in RSMA-based satellite networks.This algorithm maps the iteration process of WMMSE into a layer-wise structure similar to neural network for training.A projection gradient descent(PGD)scheme is adopted to substitute the complex operations in the original algorithm and momentum acceleration is also introduced to avoid local optimum.The simulation results show that the proposed algorithm has higher weighted sum rate and better generalization capability comparing with original WMMSE algorithm.2.An improved PSO algorithm and an improved SAC algorithm are proposedThis dissertation proposes an improved PSO algorithm and an improved SAC algorithm to optimize the energy efficiency in RSMA-based satellite networks,respectively,which can enhance the poor performance and learning ability of conventional successive convex approximation(SCA)approached.The improved PSO algorithm optimize the search process to balance the low efficiency caused and the local optimum caused by slow global search and fast local search.Additionally,also to jump out of local trap,particle resampling is introduced to judge and guide the particles movement.The improved SAC algorithm designs a trigger mechanism to avoid learning repetitious actions in the late phase when the differences of former state and current one is unobvious,which could increase the learning efficiency.This mechanism could also avoid local optimum in the early phase if actions are chosen merely based on the states change.The simulation results show that the two proposed algorithms have higher energy efficiency than conventional SCA algorithm.And they improve their performance and reduce computational cost comparing with their own original algorithms,respectively.3.An improved EVO algorithm and a constrained SAC algorithm are proposedThis dissertation proposes an improved EVO algorithm and a constrained SAC algorithm to optimize multiple rate metrics in power allocation RSMA-based multi-beam satellite networks.In multi-beam satellite networks,the adoption of multiple beams and the rapid growth of users lead to the increment of non-linearity of problem and dimension of solution space.Traditional gradient search approaches cannot deal with this situation.Meanwhile,since the conflict between network throughput and user fairness becomes more prominent in large scale networks,their balance must be considered to meet different requirements of practical networks.Since the dimension of solution space and the number of feasible solutions are rapidly growing,the performance of EVO is affected by the distribution of initial solutions.An “extreme-average” generation mechanism of initial solutions is presented for better uniform distribution.Besides,the random movement of EVO is optimized as well to avoid local optimum.The coupling of actions caused by the growth of users in multi-beam scenario will bring about difficult and slow learning process.So,the action space is modified as multi-level power allocation coefficients to decouple actions.Meanwhile,to guarantee the convergency with different metrics,the reward functions are re-defined according to their implicit constraints.The simulation results show that both algorithms could effectively optimize different metrics with strong robustness.