【作者】 王焱滨；

【导师】 虞厥邦；

【作者基本信息】 电子科技大学 ， 电路与系统， 2003， 博士

【摘要】 我们知道，在以码分多址(CDMA)技术作为首选多址接入标准的第三代(3G)蜂窝移动通信系统中，对于信道编译码、多用户检测(MUD)、软件无线电和智能天线等关键技术的研究，近年来已受到人们的广泛关注。本文主要致力于计算智能方法在多用户检测中的应用研究。由于采用最大似然检测(MLD)的最优多用户检测方法具有指数的计算复杂度，因此，研究能够有效抑制多址干扰（MAI）、具有低误码率（BER）和合理的计算复杂度、对远近问题不敏感的次优检测方法是本文的主要内容。本文主要包括以下创新之处：(1) 本文首先提出了一种基于禁忌搜索的多用户检测算法。这种方法采用传统检测器的输出作为初始解，以与当前解的汉明距为1的点组成邻域，每一次搜索所得解直接置于禁忌表中并令其永远处于禁忌状态。这种方法具有多项式的计算复杂度，对远近问题不敏感，并且能够得到良好的检测误码性能。(2) 结合禁忌搜索和多级检测，提出两种混合多用户检测算法：一种方法是将多级检测应用于禁忌搜索算法每次迭代所产生的解，另一种方法是将多级检测应用于禁忌搜索算法每次迭代邻域中的所有点。仿真结果表明，相对于单独采用多级检测或禁忌搜索的多用户检测方法，这两种方法的混合既可以减少禁忌搜索算法的计算量，又可以改善多级检测方法的性能。(3) 提出一种应用禁忌学习神经网络的多用户检测方法。这种方法将多用户检测目标函数转化成神经网络能量函数；根据禁忌搜索的概念在能量函数中引入罚项，从而解的搜索朝着未访问过的状态方向进行，这就使得搜索过程能够避免陷入局部极小值点，最后得到全局最优或近似全局最优解。这种方法具有平方的计算复杂度，仿真结果验证了其全局收敛性。(4) 利用遗传算法和Hopfield神经网络的优点，提出一种基于遗传算法和神经网络的多用户检测器。该检测器中，遗传算法首先给神经网络提供一个较好的初始解，神经网络在此基础上按梯度下降的机制进行局部寻优。这种GA和HNN结合的方法具有平方的计算复杂度，相对于单独采用遗传算法的检测器，能够极大地减少计算量；而且能够获得比单独采用Hopfield神经网络更好的检测性能。<WP=5>(5) 提出一种新的径向基函数神经网络多用户检测方法。该方法采用自适应投影算法来构造和训练径向基函数神经网络，只需一组接收信号训练样本，就可以通过迭代确定RBF函数的个数、中心的位置和网络的权系数。这种方法对远近问题不敏感，相对于有监督聚类和K平均聚类的径向基函数神经网络检测方法，节省了先验信息的需求，并能获得与有监督聚类RBF网络检测方法接近的良好检测性能。更多还原

【Abstract】 It is well known that in the next generation (3G) cellular mobile communication systems the code-division multiple access (CDMA) has become the dominant technical standard and related key techniques, such as channel coding/decoding, multiple user detection (MUD), software radio as well as intelligent antenna, are attracting increasingly research interest in recent years. This thesis is dedicated to the application of computational intelligence methods to solve the difficult issue of MUD design capable of canceling the so-called multiple access interference (MAI) to reach low bit error rate (BER) and high near-far resistant capability with acceptable computation complexity. Our attention is focusing on the sub-optimal MUD algorithm development since the maximal likelihood detection (MLD) based optimal MUD has been shown to have the exponential computation complexity. The main contribution of this thesis can be summarized as follows:(1) A tabu search (TS) based MUD algorithm is firstly proposed, in which the output of a conventional detector is taken as the initial solution, and those points whose Hamming distance to the current solution is 1 are gathered into the neighborhood, then search results of each iteration are put into tabu list and make it tabu forever. This TS-MUD algorithm is shown to be near-far resistant and of low BER with polynomial computational complexity.(2) Two hybrid algorithms by merging the TS and Multi-Stage Detection (MSD) technique are developed: 1) the MSD is used on the output solutions with respect to each iteration of the TS procedure; 2) the MSD is embedded into the TS and used onto the neighborhood at each iteration. Performance improvement of the two algorithms are observed in comparative simulation experiments with respect to the above TS-MUD algorithm. (3) A sub-optimal MUD based on an artificial neural network (ANN) with tabu learning is proposed. In this algorithm, the MUD objective function is mapped onto the energy function of the ANN, a penalty section is added to the energy function according to the TS rule, upon which any solution search always towards the states that has not been visited. This procedure enables the state trajectory to climb out of local minima thereby to converge toward the optimal or a near-optimal solution. This algorithm, justified by simulation experiments, is extremely effective due to its global convergence capability together with square computational complexity.<WP=7>(4) By taking advantages of the GA (genetic algorithm) and HNN Hopfield neural network, a hybrid MUD algorithm is presented. In this detector, GA provides firstly an initial solution at first, upon which the HNN performs local optimization according to the steepest descent mechanism. This novel hybrid algorithm, featuring also merely square computational complexity, requires a much smaller population size and generation number as compared with the detector using GA alone. This fact makes it much efficient than that of GA based MUD and HNN based MUD.(5) A novel MUD based on a radial basis function neural network (RBFNN) trained by an adaptive projective learning algorithm is proposed. Taking only a group of samples of received signal, this approach can identify the number of RBF function, the centers and the weights of the RBFNN. The proposed MUD algorithm is near-far resistant. Besides, the algorithm needs less a priori system information as compared with the detectors using K-Means Clustering RBFNN and Supervised Clustering RBFNN, but exhibits nearly the same good performance as that of the later RBFNN based detector.更多还原