Figure 1 shows the structure of the channel simulator used in this paper, which consists of three parts. Noise generator produces complex white Gaussian noise. Doppler filter filters white noise to generate correlated noise with its PSD complying with Jakes PSD. Interpolator makes Doppler rate of the fading process output satisfy the requirement of the physical channel. The Doppler rate is defined as, where is the sampling rate of fading process [5].

The value at the Doppler filter usually takes 0.2. When implemented, Doppler filter remains unchanged and interpolation factor of the interpolator is changed to adjust at the output:. This article will focus mainly on the design of the Doppler filter next.

One important consideration in designing Doppler filter is to ensure the stability. Generally, the larger the filter order is, the more difficult to keep it stable. On account of this, a good choice is to use the cascaded low order IIR sections as the Doppler filter. Except for being easy to be stable, this structure has additional advantage that is less sensitive to quantization error compared with high order IIR filter.

Formula (3) shows the system function of the traditional biquad IIR filter, where a, b, c, d are real-valued optimization variables. Reference [9] studied the stable condition of this biquad, which was called the “Triangle Stability Region” but not intuitive.

We factorize its numerator and denominator, writing it into the form of conjugate poles and zeros (4), where zr, zi, pr and pi are real-valued optimization variables, representing the real and imaginary parts of zero and pole respectively.

This Second-Order Section (SOS) is inherently stable, as long as all poles are strictly inside the unit circle. When implementing in a physical system, restore

the conjugate form into its traditional biquad form.

Usually, multiple of SOSs need to be cascaded to approximate the objective response. Assuming we use SOSs, and uniformly sample variable into points points along the unit circle: (usually M = 500), to get the frequency response of Doppler filter after frequency sampling:

wherein is the scaling factor and is a vector containing all optimization variables:

Ideal amplitude response of Doppler filter is the square root of Jakes PSD (2) but cannot be realized with a physical system according to Paley-Winner theorem, because at the frequency point the amplitude of Jakes PSD changes from into zero immediately. For this reason, [4] gives a modified objective response that can be achieved by a stable system. Referring to it herein, this article uses (8) as the objective amplitude response, where is the index of frequency point.

In the formula above, parameter SA denotes the stopband attenuation relative to the passband. Objective function could be defined as:

where is a weight vector to emphasize the error minimization for certain frequency band [7]. The complete optimization problem is:

wherein constraint is an empirical condition. limits poles of the SOS located inside the unit circle to keep Doppler filter stable. Constraint confines zeros lying inside the unit circle too, so that the Doppler filter will be a minimum phase system and have the minimum group delay.

Non-convexity of (10) confirms only if one point exists in the feasible region, around which the objective function is not convex. We will find such next.

According to the first-order conditions of convexity, is convex if and only if the feasible region () is convex and (11) holds for all [10].

Here denotes the gradient of respect to, and is the incremental step of.

For problem (10), is obviously a convex set. To show the non-con- vexity of, we randomly generate from and approximately calculate using its finite difference in computer. Then change gradually and plot and its first-order approximation to see if equation (11) is satisfied.

An example is showed in Figure 2, in which we see that is not convex around. Indeed, much more than one points like could be found in simulation, at which is not convex.

Genetic Algorithm (GA) simulates the evolution process of biotic population with selection, crossover and mutation operations. The population evolves after several generations to achieve a stable situation, which corresponds to the convergence of optimization problem [11]. Procedures in simulating GA have been described as pseudocodes in Figure 3.

Some terminologies of GA will be used in the subsequent content: A possible solution of the optimization problem is called an individual in GA. Each individual consists of a number of genes and each gene corresponds to an optimization variable. For example in (7) is called an individual, and variables A, zr, zi, pr and pi in are genes. Technical details of the pseudocodes will be given in the following.

GA parameters: Supposing that Doppler filter uses K = 7 cascaded SOSs, the referenced GA parameters are listed in.

Table 1. Generally, the larger Nind is, the greater probability that the global optimal solution may be contained in the population, so Nind should be configured big enough to keep a good diversity of the genes.

Initializing population: This step generates the initial population and each member must be feasible, meaning that individuals must satisfy constraints in (10). Firstly, generate genes A, zr, zi, pr and pi from uniform distribution in the range [−1, 1] as an individual.

Then checking the feasibility, add the individual into the initial population if its poles and zeros are all within the unit circle, otherwise discard. Repeat this process until Nind feasible individuals are created to form an initial subpopulation.

Calculating fitness: Fitness is used to quantitatively describe each individual’s adaption to the environment. This article chooses the value of objective function as fitness and the value smaller represents better adaptation.

In this step, compute the objective function value of individuals in each subpopulation first, then sort all individuals in an ascend order according to their fitness values.

Selecting elite individuals: Elite individuals refer to those who have the best adaptabilities in every subpopulation. They are preserved and passed to the next generation directly, which ensures that the performance of the new population will not degrade. This step selects Nelite individuals in each subpopulation, whose objective function values are smallest.

Selection: Selection is to select Nparent individuals from each subpopulation to produce offspring. Individuals stronger will have more opportunities to be selected and mate, while those with poor adaptability will be eliminated gradually. All individuals selected will be divided into two parts, where the first part will do crossover operation, and another part will mutate. In crossover, individuals need to be selected as parents because two parental individuals will produce one child. As for mutation, Nmut parents are needed because one single parent mutates to produce one child.

In this paper, a tournament selecting method [12] has been used. S (tournament size, e.g. 4) players are randomly chosen from each subpopulation in the tournament. A player is an individual and the one that has the best fitness value will be selected. Repeat this process until Nparent parental individuals are selected in each subpopulation.

Crossover: Crossover simulates the mating process of natural population. In the operation, parental individuals are paired first and each pair consists of a father and a mother. Then randomly select half of genes from the father, and the complementary half from the mother. Finally combine these two parts of genes together to produce a child.

The subscript is the index of gene in the individual and rnd is a random integer generated from 0 and 1. Totally Nxov children in each subpopulation are need to be produced in this step.

Mutation: In a biotical population, gene mutation happens when passing genes from parents to children, which will produce new genes and increase the genetic diversity of the population. Gaussian mutation [13] is one kind of classical mutation methods in GA, where a random variable, generated from the Gaussian distribution (16) with zero mean and standard variance, will be superimposed on the original gene as a mutation variation.

At the beginning of the GA, each gene is dispersedly distributed in the entire feasible region. The variation magnitude can be relatively big at this moment to offer more possibilities for each gene. However, as the evolution progresses, each gene begins to converge and variation magnitude should be reduced accordingly. Because after evolution, gene values that were far away from their values, now have been eliminated already by natural selection. If the variation amplitude generated is too large, the gene newly mutated will still be eliminated with great probability in the subsequent evolution, which indicates that such mutation has no meaning.

This article chooses the standard variance of to be of the Gaussian distribution. As the iteration progresses, will gradually converge, tracking this tendency becomes smaller. Thus the variation magnitude will also decrease, which avoids invalid searching.

Combination: After mutation, all offspring have been produced and they will be combined to generate a new population, which consists of three parts. The first one is the elite individuals retained from the population of the last generation. The second one is the offspring created by crossover operation and the last part is the children generated by mutation, whose genes are mutated versions of their parents’ genes.

Population newly generated will be delivered to calculate fitness again and the generation number Gen will be updated. So the evolution continues and optimization variables converge to the global or near global optimum solution gradually.

Migration: Migration occurs every MigG generations past promoting the communication between subpopulations. Elites of each subpopulation are gathered up first in this step, then they are randomly divided into Nsub parts, passed to each subpopulation to replace their worst fit individuals. Subpopulation and migration mechanism can efficiently improve the probability of getting the global optimal in GA.