1

Rate Adaptive Resource Allocation for Multiuser OFDM using NSGA – II Nitin Sharma†, Adithya Rao†, Akshat Dewan†, and Mustafa Safdari†, Nonmembers BITS, Pilani – Goa Campus, Zuarinagar, Goa (India) – 403726 Email: [email protected], [email protected], [email protected], [email protected]

Abstract-- This paper presents a new rate adaptive resource allocation technique for multiuser Orthogonal Frequency Division Multiplexing (OFDM) systems. We optimize both bit and subcarrier allocation by considering Rate maximization and Total Power constraint satisfaction. We solve them effectively by combining them into a multi-objective optimization problem. We propose using a Non Dominated Sort Genetic Algorithm (NSGAII) – a multi-objective optimization using Genetic Algorithm. Instead of combining many conflicting objectives into a single function, the NSGA-II uses multiple objective optimizations and brings out solutions which provide a better trade-off taking all conditions into consideration. The simulation results and their marked improvement over previous algorithms provide the basis to this. Index Terms – Maximum Power Utilization, Multi-Objective Optimization, Non-dominated Sort Genetic Algorithm (NSGAII), OFDM, Rate Adaptive Resource Allocation

I.

INTRODUCTION

The advent of new generation of communication technologies has ushered in an era of high data rates and better reliability. OFDM is a promising modulation technique which mitigates the effect of frequency selective fading which is inherent in high data rate environments. OFDM is used in wireless LANs, and can support high data rate transmission. It can also be used for multiple accesses [1]. Resource allocation in Multiuser OFDM [2], [3], [4], [5] includes subcarrier allocation, power allocation, and bit loading. Developing efficient resource management techniques for such a setup has drawn enormous attention in recent years. Solution to the resource allocation problem in multiuser OFDM has been broadly divided into two categories: Margin Adaptive (MA) and Rate Adaptive (RA) [4]. Resource allocation was tackled in [5] using margin-adaptive scheme, wherein an iterative subcarrier and power allocation algorithm was proposed to minimize the total transmit power given a set of fixed user data rates and the Bit Error Rate (BER) requirements. In [2], rate adaptive method was used, wherein the objective was to maximize the total data rates over all users subject to power and BER constraints. It was shown in [2] that in order to maximize the total capacity each subcarrier should be assigned to the user with best gain on it.

The MA Optimization technique has been dealt with efficiently in [6], [8]. Genetic algorithm has been used here for resource allocation and has given better results than normal iterative algorithms. But only subcarrier allocation was considered as an objective for GA and bit loading was done using Water-filling Algorithm. In [7] it was shown that RA optimization can be solved sub-optimally by separating Subcarrier and Bit Loading Allocation. We consider a combined solution to both – subcarrier allocation and bit loading using NSGA-II in the same GA which reduces complexity. Rate maximization and satisfying total power constraints are two seemingly conflicting objectives. To solve such an RA optimization, we may combine them into a single objective. But the solutions will not be optimal considering all possible trade-offs. If they are considered as two separate objectives, then the solutions will cater to most of the trade-offs. In this paper, we propose using NSGA-II to optimize RA considering two such objectives: Rate maximization and Total Power Utilization. If the Power Constraint can be relaxed and the Total Power is limited to a small range with a specified typical value, then the RA optimization problem can be solved more effectively. We use NSGA-II as it combines the problem solving capability of Genetic Algorithms without being constrained to use a single objective. II.

PROBLEM FORMULATION

Consider a multiuser OFDM system with K users and N subcarriers. The base station needs to assign a set of subcarriers to each user (subcarrier allocation) and also decide upon the number of bits per OFDM symbol to be transmitted through each subcarrier (bit allocation).Each user’s data are distributed across the set of subcarriers assigned to the user. Adaptive modulation takes place at each subcarrier based upon the number of bits assigned to it. The assumption is that each subcarrier is uniquely assigned to a single user and two or more users never share the same subcarrier. The optimization problem is formulated as described in Wong [7] and the notation used there is adopted throughout. Let Rk be the data rate of the kth user. Suppose the nth subcarrier is assigned to the kth user. Let the number of bits of the kth user being transmitted through the nth subcarrier be

2 denoted by ck,n where ck,n is less than or equal to M, the maximum bits/symbol that can be transmitted by each subcarrier. The data rate can be expressed as N

(1)

n =1

We can also define an indicator variable

ρ k , n as

0, if ck , n = 0 1, otherwise

ρ k ,n = 

(3)

Since each subcarrier is occupied by a single user,

=1

(4)

k =1

α k2,n is the channel gain for the nth subcarrier of the kth user

then the transmission power assigned to it is given by

Pk ,n =

f k (ck ,n )

(5)

α k2,n

where f k (ck , n ) is the power required for reliable reception of

ck ,n bits/symbol for unit channel gain. For QAM, f k (c ) can be expressed as 2

N   p  f k (c) = 0 Q −1  e   (2c − 1) (6) 3   4  where pe is the Bit Error Rate (BER) and N 0 / 2 is the variance of the Additive White Gaussian Noise (AWGN), and Q ( x ) is the Q -function . With these notations the RA optimization can be formulated as follows: Formulation: The RA procedure described in Kim [7] maximizes the minimum data rate among all the users subject to the constraint that the total power cannot exceed a given value. Here we modify the power constraint slightly and assume that the total available transmission power is limited to a certain range with a typical value PT .We then have a second objective-- to bring the total power as close to PT as possible. The new multi-objective optimization problem is

N

min PT − ∑∑

ck ,n , ρ k ,n

k =1 n =1

k

n =1

f k ( ck , n )

α k2, n

ρ k ,n

(7b)

III. ALLOCATION USING NSGA-II

n =1

If

ck ,n , ρ k ,n

(2)

N

Rk = ∑ ck ,n ρ k ,n

k ,n

k

subject to the constraint (4).

so that

∑ρ

ck ,n , ρ k ,n

K

Rk = ∑ ck ,n

K

N

max min Rk = max min ∑ ck ,n ρ k ,n (7a)

Genetic Algorithms: Genetic Algorithms (GAs) [6], [13], [14] are a class of evolutionary algorithms. They provide novel approaches to problem solving technique inspired by biological evolution. GAs enable efficient search in the solution space of any function so as to get a solution set that optimizes the function. This efficiency with GAs is due to operators analogous to the ones found in natural evolution of species (to evolve better species): Selection, Crossover, and Mutation. Every solution in the population is evaluated by a “fitness function” and assigned a “fitness” value which indicates how favorable the various traits of that solution are, and how much the solution optimizes the fitness function. This fitness value decides the participation of a solution in evolving an optimum solution. Through crossover, fragments of the two different solutions are mixed to give rise to new offspring with combined traits of both parent solutions. Through mutation, a wider set of the actual solution space is explored (more than what is available in the initial population). Every intermediate solution set (population of chromosomes), called a generation, goes through fitness evaluation, crossover and mutation to evolve fit solutions. The basic GA works to optimize a single objective function. However, many optimizations and resource allocations involve tradeoffs between various objectives and parameters and therefore are multi-objective [11], [12]. For two conflicting objectives, each objective corresponds to a different optimal solution. Thus in multi-objective problems, there is no single optimum solution, but many acceptable solutions. The emphasis then shifts to finding a solution that – handling minimum conflicts – delivers a satisfactory solution for all objectives. A higher level qualitative choice can be made for one possible solution among the set of optimum solutions. NSGA–II: Non-dominated Sort Genetic Algorithm (NSGA) uses an effective non-domination sorting algorithm to optimize multi-objective functions [18]. NSGA-II improves the computational complexity of NSGA and also incorporates elitism [9]. NSGA solves multi-objective problems by using the concept of domination. In NSGA-II the initial population is sorted into fronts [9]; where the individuals in the first front are not dominated by any other individuals in the current

3 population, the individuals in the second front are dominated only by the individuals in the first front and so on. An individual solution is said to dominate another if its fitness values with respect to every objective fitness function is superior to the corresponding values for the other individual. The individuals in the rth front are assigned a rank of r. In addition, crowding distance, a parameter that measures how close an individual is to its neighbors, is also calculated for each individual. A larger average crowding distance indicates a greater diversity in the population. In [10] the MA optimization problem was addressed using GA, where each individual chromosome was coded as an array of elements that represented subcarriers. Here we extend the analogy to simultaneously allocate subcarriers and bits, while taking constraint (4) into consideration. Each individual chromosome is coded as an array of N elements, where N is the number of subcarriers. Each of these N elements is binary coded with (log2K+log2M) bits, where the first (log2K) bits of the nth element represent the user to which the nth subcarrier is assigned, and the next (log2M) bits represent the number of bits allocated to the nth subcarrier. The size of each individual chromosome is given by N ×(log 2 K + log 2 M ) . Since each subcarrier is assigned to exactly one user, constraint (4) is satisfied. The last row of Fig.1 depicts how the chromosome would be considered for crossover and mutation. The steps involved in NSGA-II are as follows [9], [11]: 1. Population Initialization: The number of individuals in the population, P and the number of generations, G are fixed beforehand and can be changed for different runs of the algorithm. The 2-D population consists of bits. The Population size (for some K, M, N and P) is P × ( N × (log 2 K + log 2 M )) . Each individual is created by generating a random string of 0’s and 1’s. 2. Evaluate Objective functions: The fitness values of the objective functions (7a) and (7b) are calculated for each individual. 3. Non-Dominated Sort: The population is now sorted into fronts based on non-domination and each individual is assigned a rank. After sorting, each individual is assigned crowding distance front-wise. Crowding distance is based on finding the Euclidean distance between individuals of a particular front. The individuals at the boundaries are assigned infinite crowding distance. 4. Tournament selection: A binary tournament selection procedure is used to fill the mating pool, which is taken to be of size P/2.Selection is based on rank, and if two individuals have the same rank then the selection is based on crowding distance. An individual with a lower rank and a higher crowding distance is more favorable.

SC1 SC2 K1 C1 K2 001 01 011 0010101110.... 11001

..... .....

C2 10

SCN KN 110

CN 01

Fig.1 Schematic representation of a chromosome.

5.

Crossover and Mutation: Offspring chromosomes are created by performing the genetic operations of crossover and mutation on the parent chromosomes in the mating pool and their fitness values are evaluated. 6. Generation of new population: An intermediate population consisting of the parents and offspring of the current population is created and is sorted using non-domination. The new population is filled up by taking the best individuals from the combined population based on rank and crowding distance. 7. Steps 4, 5 and 6 are repeated for G number of generations. In [8] elitism was ensured by using Largest Weighted Delay First, LWDF. Here since all the best individuals from the current and previous populations are added to the new population, elitism is guaranteed. The best individual chromosome from the final population gives the desired allocation of subcarriers and bits per subcarrier. IV. S IMULATION AND RESULTS Assumptions and Constants: For various constants required in the calculation, their accepted values were used: BER = 10-7, N0 = 10-8, M = 8 The channel gain

α k2,n for

kth user on the nth subcarrier is

generated randomly using Normal Distribution N (0, 1). This channel gain matrix is generated once and kept constant throughout the simulation. Total Power (PT in the constraint) has been varied across simulations from -10dB to -40dB assuming the reference power level for dB calculations is 1. For every run of the algorithm we have considered permutations of the number of subcarriers N, and number of users K (8) K ∈ [2, 64]; N ∈ [2,128]; K ≤ N ; K , N ∈ Z TABLE I R ESULTS ON RUNNING THE ALGORITHM FOR VARIOUS PERMUTATIONS OF NO . OF USERS, NO . OF SUBCARRIERS AND POPULATION SIZE Pc,Pm Sr. Users Subc. Optimized min. Convergence No. (K) (N) Rate (gen.) pop=30

pop=50

pop=30

pop=50

1

4

16

0.6,0.03

7.8

11.8

25

25

2.

8

128

0.9,0.03

52.9

44.4

25

25

3.

16

64

0.9,0.03

5.6

6.4

35

30

4.

2

16

0.9,0.03

30.3

40.2

25

25

5.

16

16

0.85,0.15

0

0

*

*

* Premature Convergence from very early stage

4 The results (Fig.2, Table.1) obtained for the various simulations reaffirmed that applying multi-objective optimisation to resource allocation problems in multiuser OFDM gives more efficient solutions. Firstly, the number of generations required for the GA to converge was much less than previous algorithms (60 generations in [6] compared to 30 in this method in best case). Secondly, both the objective functions converged despite them being of conflicting nature (Fig.2). Thirdly, the results were consistent with the expected results for K < N. For large number of subcarriers, the optimized Rate is high (Fig.3) but within bounds as warranted

by the Total Power constraint. We also observed that if we could estimate the maximum typical Total Power (< PT) that would be utilized, based on the channel conditions and the data rate requirements, and if we set our constraint near to this value, the algorithm converges much faster than without this. It was also found that as the population size increased, there was a marked improvement in the performance of the GA, as is evident in Fig.3 and the algorithm produced a better result. This is consistent with the expected performance of efficient GAs.

Fig.2 Simulation results for the first 3 permutations of conditions as described in Table I. The graph on the left side depicts the Rate optimization which is minimizing (1/ (1+RMIN)). The graph on the right side depicts optimizing the Power Constraint which is minimizing the difference between the typical Power and the Power consumed by the current allocation scheme of the best individual.

5

Fig.3 Graph showing performance of the system with increasing no. of users

In the case of K=N (case 5 in Table 1), the randomly generated initial population causes the algorithm to be stuck at the local minimum and tends to make RMIN=0. It can thus be inferred that the algorithm works best only for K
G × [O( PKN ) + O( JP 2 ) + O( J log P) + O( P)] 2 ≅ O(( PG )( KN + JP ))

(9)

where

G = Number of Generations, P = Population Size K = Number of Users, N = Number of Subcarriers J = Number of Objectives = 2 For some constant N, the complexity of the algorithms is O(GP2K) which is comparable to that suggested in [8]. TABLE II C OMPLEXITIES OF VARIOUS OPERATIONS IN THE ALGORITHM

Operation

Complexity

Evaluating fitness function for each individual

O(PKN)

Non-Dominated Sorting

O(JP2)

Crowding Distance Calculation

O(J log2P)

Tournament Selection

O(P)

V. CONCLUSIONS In this paper, we have proposed the use of NSGA-II for resource allocation in multiuser OFDM, which depends on inherently conflicting multiple objectives. The results produced by the simulations indicate that the algorithm converges in lesser number of generations compared to other implementations of GA [6]. The complexity of this algorithm is comparable to the complexity obtained in [8], showing that the addition of another objective does not add complexity or computational overhead to the algorithm. The solution generated by the algorithm is acceptable considering both objectives. We have also proposed a method for faster convergence by reformulating the constraint of total power. In addition, both subcarrier allocation and bit loading have been optimized using NSGA-II. In [8] it was shown that using diversity to adapt Genetic Operator values gives better results. NSGA-II calculates crowding distance as a measure of diversity, and this calculation is a part of the algorithm and not an overhead operation. Therefore, the adaptability of operators based on diversity can be easily implemented using the crowding distance. This can be a possible extension of this technique in the future. REFERENCES [1]

David Tse and Pramod Viswanath, Fundamentals of Wireless Communication Cambridge University Press 2005 [2] J. Jang and K. B. Lee, “Transmit Power Adaptation for Multiuser OFDM System” IEEE J. Select. Areas Commun, Vol. 21, pp. 171 to 178, February 2003 [3] Zukang Shen, Jeffrey G. Andrews and Brian L. Evans, “Optimal Power Allocation in Multiuser OFDM Systems”, Proc. IEEE Global Commun. Conf., San Francisco, CA, Dec. 2003, pp. 337-341 [4] Ian C. Wong, Zukang Shen, Brian L. Evans and Jeffrey G. Andrews, “A Low Complexity Algorithm for Proportional Resource Allocation in OFDMA Systems” Proc. IEEE Int. Workshop on Signal Processing Systems, pp. 1-6, Oct. 2004. [5] C. Y. Wong, R. S. Cheng, K. B. Lataief and R. D. Murch, “Multiuser OFDM System with Adaptive Subcarrier, Bit and Power Allocation” IEEE J. Select. Areas Commun, Vol. 17, pp. 1747 to 1758, October 1999 [6] Y. B. Reddy, “Genetic Algorithm Approach for Adaptive Subcarrier, Bit, and Power Allocation” Proceedings of the 2007 IEEE International Conference in Networking, Sensing and Control, London, UK, 15-17 April 2007 [7] Inhyoung Kim, In-Soon Park, and Yong H. Lee, “Use of Linear Programming for Dynamic Subcarrier and Bit Allocation in Multiuser OFDM” IEEE Transactions on Vehicular Technology, Vol. 55, No.4, July 2006 [8] Zhihua Tang, Youtuan Zhu, Guo Wei, Jinkang Zhu, “Cross-Layer Resource Allocation for Multiuser OFDM Systems based on ESGA” Proceedings of the 66th IEEE Vehicular Technology Conference (VTC 2007 – Fall), Baltimore, MD, USA, October 2007 [9] Kalyanmoy Deb, Amrit Pratap, Sameer Agawal, and T.Meyarivan “A Fast and Elitist Multi-Objective Genetic Algorithm: NSGA-II”, KanGAL Report Number 2000001 [10] Yenumala B. Reddy, Nandigam Gajendar, “Computationally Efficient Resource Allocation in OFDM Systems: Genetic Algorithm Approach” Proceedings of the 4th International Conference on Information Technology: New Generations (ITNG 2007), Las Vegas, Nevada, USA, April 2-4 2007, [11] Kalyanmoy Deb, Multi-Objective Optimization using Evolutionary Algorithms, Wiley, 1 edition (June 27, 2001)

6 [12] Lance Chamber Ed. Practical Handbook of Genetic Algorithms Volume 1, CRC Press 1995 [13] David E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Pearson Education 2006 [14] Earl Cox, Fuzzy Modeling and Genetic Algorithms for Data Mining and Exploration, Elsevier 2005 [15] Ian C. Wong and Brian L. Evans, “Optimal Downlink OFDMA Subcarrier, Rate and Power Allocation with Linear Complexity to Maximize Ergodic Weighted-Sum Rates”, Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, Honolulu, HI, April 2007 [16] Ian C. Wong and Brian L. Evans, “Joint Channel Estimation and Prediction for OFDM Systems”, Proc. IEEE Global Telecommun. Conf., St. Louis, MO, Dec. 2005, pp. 2255-2259 [17] Ian C. Wong, Antonio Forenza, Robert W. Health and Brian L. Evans, “Long Range Channel Prediction for Adaptive OFDM Systems”, Proc. IEEE Asilomar Conf. on Signals, Sys. and Comp., Vol.1, CA, Nov. 2004, pp 732-736 [18] N. Srinivas and K. Deb, “Multi-Objective Function Optimization using Non-Dominated Sorting Genetic Algorithms”, Evolutionary Computation Vol. 2, pp 221 – 248, 1995