Orthogonal frequency-division multiple-access (OFDMA) has emerged as a promising multiple-access technique for next-generation wireless networks operating over wideband multi-path fading channels. Classical OFDMA, which involves assigning orthogonal (non-overlapping) subcarriers among different users, is an effective means to handle multi-path fading with low-complexity receivers and to obtain multi-user diversity gains via channel-dependent frequency-domain scheduling. Consequently, resource allocation for classical single-cell OFDMA systems has been the subject of intense research in recent years. The methods designed for single-cell OFDMA systems are directly applicable to a multi-cell OFDMA system employing static frequency reuse wherein inter-cell interference is avoided by ensuring adjacent cells are allocated orthogonal frequency bands. This allows for autonomous resource allocation at each base-station at the expense of reduced system throughput. The resource allocation problems for OFDMA systems are in general non-linear and non-convex which has prompted several heuristic approaches. It is also known that many OFDMA resource allocation problems satisfy a time sharing property in the limit of large number of subcarriers and hence in principle optimal solutions can be obtained via dual decomposition techniques. Unfortunately, the complexities of the dual decomposition technique as well as many heuristic algorithms are still too high for most practical cellular systems being considered for deployment.
Future wireless network evolutions are envisioned to employ a full (or an aggressive) frequency reuse. However, co-channel interference caused by transmissions in neighboring cells will be a major impairment that limits throughput. Consequently, proactive inter-cell interference mitigation techniques are required. Joint resource allocation over a cluster of neighboring cells via base-station coordination is a promising solution. However, most recent works focused on downlink systems and the design of distributed uplink resource allocation algorithms that require limited information exchange among base-stations is still in its infancy.
Systems and methods for performing radio resource allocation system based on the auction method for uplink OFDMA cellular networks are disclosed.
In one aspect, Systems and methods to assign one or more resources in a multi-user single cell Orthogonal Frequency-Division Multiple Access (OFDMA) uplink includes specifying a resource allocation problem for the one or more resources; converting the resource allocation problem into an assignment problem; solving the assignment problem through an auction; and allocating the one or more resources to cellular users to maximize a system utility.
Advantages of embodiments of the system may include one or more of the following. The system allocates resource based on the auction method with very low complexity and can offer near-optimal performance. The system works with cellular systems that employ the traditional static frequency reuse as well as the next-generation systems that aim to achieve a universal frequency reuse via base-station coordination. The system works with finite input alphabets and also accounts for non-ideal practical outer codes, and can be implemented in a distributed manner, when applied for multi-cell resource allocations. The system also handles power and bandwidth constraints that are motivated by practical concerns. The system exhibits very low complexity and is well suited for parallel implementations. Simulation results demonstrate that it offers near-optimal performance.
The system provides efficient uplink resource allocation for OFDMA networks. Here the term resource refers to subcarriers, power as well as the modulation and coding schemes (MCS). The resource allocation processes provide practical MCS rather than the commonly assumed Gaussian input alphabets and ideal outer codes. While Gaussian alphabets greatly simplify analysis, their main drawback is that the capacity with Gaussian input alphabet is unbounded in the signal-to-noise ratio (SNR) which results in too much power being allocated by a user to its stronger subcarriers. As a remedy many works introduce an SNR gap (or penalty). The system explicitly considers practical modulations such as QAM in the context of a multi-user OFDMA network and accounts for non-ideal (practical) outer codes. The system also reduces signaling overhead through strategies that minimize the amount of information the base-station has to feed back to the users. A signaling reduction technique is used that allows only equal (or constant) power allocation for each active user, i.e., each user must transmit with the same power on all its assigned tones. The system imposes bandwidth constraints that are motivated by practical concerns such as capabilities of mobile terminals as well as the fact that the cell-edge users may be able to transmit pilots with enough power only on a few subcarriers and consequently they may be scheduled only on those subcarriers.
In a low mobility scenario, the channel coherence time is large enough to accommodate feedback delays as well as sufficiently large code lengths. If N is the total number of tones, the demodulated signal on the nth tone at the BS is given by
where √{square root over (hkn)} is the positive channel gain seen by user k on tone n, vn is the zero-mean unit-variance complex proper Gaussian ((0,1)) additive noise variable. xkn ε Skn is channel input from user k at tone n and Skn is a unit average energy constellation. Moreover, Skn can be any constellation from a set of constellations Sk. Thus user k is allowed to use a possibly different constellation for each tone.
is the total power transmitted by user k in one OFDM symbol duration and must satisfy
where Pk is the specified power budget of user k. Due to the orthogonality constraint in an OFDMA system, i.e., pknpjn =0, k≠j, ∀ n the summation in (1) actually contains only one term.
Next, one exemplary modeling of a Practical MCS is discussed. In this discussion, a user k has a power and constellation assignment as {pkn,Skn}. Then,
where I(Skn, pknhkn) denotes the (coded-modulation) mutual information, is the rate it can achieve over the parallel channel model, using an ideal code of coding rate
constructed on the assigned input alphabets. In order to account for the non-ideality of practical codes, the mutual information exponential system mapping (MIESM) method is used. The MIESM method is based on the observation that for good practical codes, the block error probability (BLER) versus normalized mutual information (also referred to as the average mutual information per-bit) curve is (almost) independent of the underlying modulations and depends mainly on the coding rate. In particular, over the parallel channel model, the system can achieve an acceptable BLER (less than 0.01) with a good practical code of coding rate
where the constant Γ captures the loss due to the practical codes, which results in a sum rate of
A dummy constellation φ in Sk is used with the constraint that if Skn=φ then user k does not transmit on nth subcarrier. Also, I(φ, pknhkn)=0, ∀ pkn,
As a result, for a given power profile the maximum rate that can be achieved by user k using the available modulations and practical codes, is given by
For an ideal code Γ=0 while Γ≈0.1 for the codes used in 3GPP and this value is used to model non-ideal codes in the sequel.
Next, the rate loss due to suboptimal power allocation is discussed. The system bounds the potential performance degradation caused by sub-optimal power allocation. In this process, 1≦k≦K is the set of tones allocated to user k and {Skn}nε
is the constellation assignment, and the optimal power allocation for each user is then given by a mercury water filling (MWF) method. The MWF method determines the optimal power allocation over the parallel channel model for a given choice of input constellations. As opposed to the traditional water-filling which assumes Gaussian alphabets, the MWF explicitly accounts for the mutual information of the finite input alphabets which unlike the Gaussian alphabet case, saturates with increasing SNR. For any sub-optimal power allocation {pkn} that satisfies the power constraint with equality for each user, the achievable rate is at-most Δ bits away from the optimal rate achieved via the per-user MWF, where Δ is given by
where MMSE(Skn, hknpkn) denotes the minimum mean-square error in an AWGN channel with SNR hknpkn and input alphabet Skn.
For a user k with {pkn} be a valid allocation that satisfies the power constraint with equality, i.e.,
The optimal set of powers {{circumflex over (p)}kn} is obtained as the solution to
The optimization in (4) is a concave maximization problem since I(Skn, hknpkn) is concave in the power pkn. Also note that Slater's condition is satisfied with any Pk>0. Thus the strong duality property holds for (4). Define the Lagrangian
L({pkn,μkn}, λk)=ΣnεJ
and consider the dual function
Let g({{circumflex over (μ)}kn}, {circumflex over (λ)}k)=min{μ
g({{circumflex over (μ)}kn}, {circumflex over (λ)}k)=ΣnεJ
for any {{tilde over (μ)}kn≧0}, {tilde over (λ)}k≧0, the variables {{tilde over (p)}kn≧0} which satisfy
hknMMSE(Skn, hkn{tilde over (p)}kn)={tilde over (λ)}k−{tilde over (μ)}kn, ∀ n, (8)
must be the maximizers in (6) for μkn={tilde over (μ)}kn and λk={tilde over (λ)}k so that from (7)
g({{tilde over (μ)}kn}, {tilde over (λ)}k)=L({{tilde over (p)}kn, {tilde over (μ)}kn}, {tilde over (λ)}k)≧g({{circumflex over (μ)}kn}, {circumflex over (λ)}k)=ΣnεJ
For the given set of powers {pkn} if
then (8) is satisfied. Further using the fact that
with (5) and (9):
Note that for Gaussian alphabets (i.e., when each Skn is a complex proper Gaussian random variable) the bound can be reduced. Adding up Δk for all users, the desired result in (3) is obtained.
Next, two resource allocation problems are discussed. The resource allocation is performed on a frame-by-frame basis and all channel coefficients are assumed to be perfectly estimated by the BS at the beginning of each frame.
Type-1 Formulation: Fixed Power Per User Per Subcarrier
For an uplink OFDMA system with N subcarriers and K users., αki denotes an indicator variable such that it equals to one if the ith subcarrier is allocated to the kth user. d(Ski, pkihki)I(Ski, pkihki)−Γ log|Ski| denotes the effective rate of the kth user on the ith subcarrier when the constellation (or modulation) Ski is employed. The number of subcarriers allocated to the kth user is predetermined as nk. The power level at each subcarrier for this user is then
cki is the corresponding maximum achievable rate for the available modulations and practical codes, i.e.,
Further, if wk is the weighting factor (or priority) of the kth user, the resource allocation problem is given by
max Σk wk(Σi ckiαki)
s.t. Σk αki≦1, ∀i; Σi αki=nk, ∀k; αki ε {0,1}. (11)
In (11)
For simplicity, this type of subcarrier assignment is referred to as the type-1 resource allocation hereafter.
If a power level per subcarrier pk as well as the minimum and maximum bandwidths (in number of subcarriers) ñkl, ñku respectively, have been specified for user k, for all 1≦k≦K such that ñkupk≦Pk and
ñkl≦N. Let {tilde over (c)}ki maxS
max Σk wk(Σi {tilde over (c)}ki αki)
s.t. Σk αki≦1, ∀i, ñkl≦Σi αki≦ñku, ∀k; αki ε {0,1}. (12)
As shown in the sequel, (12) can be converted into (11) and hence with simplification, will also be referred to as type-1 resource allocation. The formulation in (12) can also accommodate another practically important constraint which imposes that a user k may be scheduled only on subcarriers in a specified set ⊂ {1, . . . , N}. This constraint is applicable to cell-edge users who may be able sound (i.e., send pilots on) only a few subcarriers with sufficient power. For such a user, {tilde over (c)}ki=−∞, ∀ i ∉
which will ensure that the optimal solution to (12) will not assign user k to any subcarrier not in
The process of
Type-2 Formulation: Fixed Total Power Per User
In the second formulation, a total power budget Pk is assigned to each user k, and the number of subcarriers allocated to user k lies between nkl and nku, where 0≦nkl≦nku≦N and
Pk is evenly split among all subcarriers assigned to the user.
Let cki n be the maximum achievable rate under the available MCS when the ith subcarrier is allocated to the kth user and the number of subcarriers allocated to the kth user is n, i.e.,
with cki00. αkin is an indicator variable such that αkin=1 if the ith subcarrier is allocated to the kth user and the total number of subcarriers allocated to the kth user is n. Likewise, βkn is an indicator variable such that βkn=1 if the number of subcarriers allocated to the kth user is n. The subcarrier allocation problem can then be formulated as
Problem (13) will be referred to as the type-2 resource allocation hereafter.
Auction Methods for Single-Cell Resource Allocation
The symmetric assignment problem is a special 0-1 combinatorial optimization problem that can be solved efficiently. For N persons and N objects, a reward fkj can be associated with assigning the jth object to the kth person. The goal of the assignment problem is to assign each object to a person such that the total reward is maximized. In particular:
max Σk Σj fkjηkj
s.t. Σk ηkj=1, ∀j; Σj ηkj =1, ∀k, ηkj ε {0,1}, (14)
where ηkj is a 0-1 indicator variable such that it is 1 if the jth object is assigned to the kth user. Unlike most combinatorial optimization problems, the assignment problem can be solved efficiently. Specifically, the LP polytope of an assignment problem is exactly the convex hull of its integer solutions, thus the optimal solution can be obtained by directly solving its LP relaxation. There exist a number of methods to solve the assignment problem, ranging from the classical Hungarian method to the primal-dual method, but they have high computational complexity and memory requirement. The auction method used herein has a low complexity and memory requirement, and can solve practical large-scale assignment problems.
The auction process is an intuitive method for solving the assignment problem. It mimics a competitive bidding process in which unassigned persons raise their prices and bid for objects simultaneously. After getting bids from all persons, objects are assigned to the highest bidder. More specifically, given a positive scalar ε and a set of prices {rj, j=1, . . . ,N} a person k is termed as happy with the assignment of object jk if the profit (i.e., reward minus price) of assigning the object jk to himself is within ε from the optimal, i.e.,
The auction method gradually proceeds towards a full assignment of objects to persons such that all persons are happy. The basic auction process is given as follows.
Process 1 Basic Auction for Symmetric Assignment Problem
The above process is guaranteed to converge to an assignment scheme that yields a total reward within Nε from the optimal objective function value of (14). The number of iterations to converge to an assignment scheme in the above auction method typically decreases linearly with increasing of ε. A proper ε can be selected to trade off the performance and number of iterations in practice. Also, the complexity of the auction process is linear at each iteration, and the memory required to store the price variables is also very low.
Single-cell Resource Allocation
Type-1 Resource Allocation Process
Type-1 resource allocation problem (11) uses the auction method. Additionally, a dummy user is used with index k=0, such that c0i=0, ∀ i and set its bandwith as
with the understanding that any subcarrier assigned to the dummy user is not used for data transmission. Without loss of generality, each subcarrier will be assigned to some user, dummy or otherwise. The process divides the kth user into nk “sub-users”. Define αkli, l=1,2, . . . ,nk, as an indicator function such that αkli=1 if the ith subcarrier is occupied by the lth sub-user of the kth user. Furthermore, let wkl wk and ckli
cki, l=1,2, . . . , nk, ∀k. Thus problem (11) is converted to
max Σk Σl wkl(Σi ckliαkli)
s.t. Σk Σl αkli=1, ∀i; Σiαkli=1, ∀k,l; αkli ε {0,1}, (16)
which is a standard (i.e., symmetric) assignment problem, and can be efficiently solved by the standard auction method, i.e., Process 1.
Eq. 12 can also be optimally solved using the auction process by introducing auxiliary variables. For convenience, the weighted rate (reward) of any user on any subcarrier is assumed to be non-negative. If
without loss of optimality, each inequality constraint Σi αki≦ñku can be replaced with the equality constraint Σi αki=ñku and the ignore the superfluous constraint Σi αki≧ñkl. Consequently, problem (11) is obtained and can be solved using the aforementioned technique. Next if
“dummy” subcarriers are introduced with the understanding that no user will use any dummy subcarrier assigned to it for data transmission. Further, the kth user can be split into ñku sub-users having identical rewards as the kth user for each one of the N (real) subcarriers. The key step now is to select any ñkl out of the ñku sub-users and set the reward of each one of these sub-users to be −∞ for each dummy subcarrier. Each one of the remaining ñku−ñkl sub-users gets a reward of zero for any dummy subcarrier. This step results in no loss of optimality and also ensures that at-least ñkl sub-users corresponding to user k are allocated real (non-dummy) subcarriers and this becomes problem (11) which can be solved optimally using the auction process.
Type-2 Resource Allocation Process
The auction method, however, cannot be directly used to solve the type-2 resource allocation problem (13). This is because the number of subcarriers allocated to each user is unknown, and therefore it cannot be converted directly into an assignment problem. A sub-optimal process for this problem is discussed next that leverages the auction method. This process consists of two stages, i.e., initial solution calculation and iterative improvement. At the first stage, a feasible solution is obtained by solving a corresponding type-1 resource allocation problem. This initial solution can achieve a weighted sum rate within a constant fraction of the optimal solution to (13). A tree-search method is used to iteratively improve the solution.
If each user must occupy at least one subcarrier, i.e. nkl≧1, ∀ k, and Q and {{tilde over (α)}ki, ∀k, i} denote respectively the optimal objective function value and the solution to the problem given below,
Eq. (17) can be obtained from (12) by replacing {tilde over (c)}ki with ckin Σk wk Σi ckin
of the optimal objective function value, i.e.,
The first claim follows immediately from the fact that for all n: nkl≦n≦nku,
where any modulation Ski, d(Ski, p hki) is monotonic in p.
As to the second result, note that Q=Σk wkΣi cki n
Further, if an ideal code with Γ=0 is assumed:
The inequality in the above formula follows from the fact that the mutual information function under any finite modulation scheme is concave in its input power level and its value is zero when the input power is zero, so that,
Upon getting an initial solution via solving (17), the system can iteratively improve the obtained solution through a tree-search type method. Each node in this tree corresponds to a set specifying the number of subcarriers assigned to each user, i.e., {nk}k=1K, and it is associated with an objective function value which can be obtained by solving the corresponding type-1 resource allocation problem given in (11) using Process 1. The obtained subcarrier assignment scheme is then used to generate a group of child nodes from the current node, and a simple method is used to calculate a lower bound to the objective function value of each child node. If the lower bound to the objective function value of a child node is larger than the objective function value of the current parent node, this child node is kept in the tree for further investigation. Otherwise, it is removed from the tree. The process is summarized below.
Process 2 Auction for Type-2 Resource Allocation
Process 2 can also initialize the queue with multiple sets. Moreover, to limit complexity the process can run for only for a specified number of iterations. Let Qk denote the best objective function value up to the kth iteration of Process 2. The system can bound the performance loss between the obtained solution and the optimal solution, i.e., Q*−Qk≦Q−Qk. In practice, if this upper bound is small, it is guaranteed that the obtained solution is close to the optimal.
The derived bound can give an accurate estimation of the real sub-optimality gap.
The first step in Process 2 provides a feasible solution which yields a sum rate within a fraction of the optimal objective function value of (13). This initialization method, however, becomes inapplicable if there are no minimum bandwidth constraints, i.e., nkl=0, for some users. In the single-user scenario without any bandwidth constraints and under the equal power constraint, the kth user will select the subcarriers with the mk-largest channel gains, where mk depends on the available maximum power, modulation set and channel gains and can be determined via the single-user CWF. It is intuitive that for most of the cases in the multi-user scenario with proportional fair scheduling, the optimal solution will not allocate more than mk subcarriers to user k. This is because due to multiuser diversity, a user is likely to be scheduled on its best (or good) subcarriers and it is unlikely to occupy more than mk of these. This motivates the following initialization method for Process 2.
Process 3 Initialization for Type-2 Resource Allocation if there are no Minimum Bandwidth Constraints for Some Users
Another procedure for sub-optimally solving the type-2 problem in (13) when nkl=0, nku=N for all k, is given in
Multi-cell Resource Allocation Process
In a multi-cell OFDMA system, a a mobile user not only competes for resources with users in its own cell, but also generates interference for base-stations in adjacent cells. A possible strategy under such scenario is to simply treat interference from other cells as noise, which may severely reduce the system capacity due to the strong interference from edge users in other cells. A better alternative is to jointly consider resource allocation across different cells, and thereby maximize the total system throughput.
Users are divided into two groups, i.e., interior users and edge users. An interior user does not generate interference for the base-stations in adjacent cells; while an edge user can cause such interference. Notice there exist various criteria to categorize users into the two groups. For example, a straightforward method is to categorize users according to their distances to the corresponding base stations. If a user is close enough to the base station in an adjacent cell, it is treated as an edge user. Another possible strategy is to divide users based on their instantaneous channel conditions. Next, “virtual” users are defined to capture the interaction of users in different cells. Every edge user generates a virtual user for each adjacent cell to which that edge user will generate interference. Every virtual user occupies the same subcarriers as its corresponding edge user and each cell must allocate non-overlapping subcarriers to its interior, edge and virtual users. In this way the system can decompose a multi-cell resource allocation problem into a group of single-cell resource allocation problems in which the interference between users in different cells are represented by constraints between edges users and corresponding virtual users. The general problem of resource allocation in multi-cell OFDMA systems is then given by
maximize: Weighted total sum rate
subject to: Constraints for interior users, edge users and virtual users in every single-cell;
Every edge user shares the same resource as its corresponding virtual users (24).
Assume there are B base stations. kj represents a user in the jth cell and represents a set of users not in cell j such that every user ki,i≠j ε
can generate interference to the jth cell. {circumflex over (k)}ij denotes the virtual user generated for the jth cell by edge user ki. The first multi-cell subcarrier allocation problem is formulated and assumes that a power level along with minimum and maximum bandwidths have been specified for each user. This assumption will satisfied in LTE cellular systems which will employ fractional frequency reuse (FFR), where a higher power level will be specified for each edge user and a lower one for each interior user. Then, let ckjl denote the rate obtained by allocating the lth subcarrier in the jth cell to user kj, nkjl and nkju represent the minimum and maximum number of subcarriers that can be occupied by user kj, respectively. αkjl is an indicator variable such that αkjl=1 if subcarrier l is assigned to user kj. The following formulation is the extension of (12) to multiple cells.
max Σj Σkj wkj (Σl ckjlαkjl)
such that Σkj αkjl+Σkjεα{circumflex over (k)}ijl≦1, ∀l,j; nkjl≦Σl αkjl≦nkju, ∀kj, j;
α{circumflex over (k)}j; αkjl ε {0,1}, ∀l,kj,j. (25)
The first two groups of constraints are the same as in the single-cell case. The third group of constraints models the effect of inter-cell interference. This type of resource allocation problem is defined as multi-cell type-1 resource allocation problem hereafter. Note that the problem in (25) is feasible if and only if N≧Σkj nkjl+Σknk
Turning now to
The solution to the LP relaxation of the problem in (25), i.e., with the relaxed constraints αkjl ε [0,1], ∀ kj,l,j, may contain fractional values for some {αkjl} and therefore may not be a feasible solution to (25). However, this solution can become feasible and hence optimal in the limit of large number of subcarriers, as shown next. The total number of users in the system remains fixed.
An optimal solution of the LP relaxation to (25) is also an optimal solution to the multi-cell type-1 resource allocation problem if the number of subcarriers in the system increases to infinity. The LP relaxation of (25) with continuous frequency variables given by
max Σj Σkj wkj(∫ ckj(f)αkj(f)df)
s.t. Σkj αkj(f)+Σkjεα{circumflex over (k)}
α{circumflex over (k)}j; αk
where δkj and γkj are the minimum and maximum bandwidths that user kj can occupy, respectively, and the integrals are over a closed bounded interval corresponding to the allocated system bandwidth. In the following an optimal solution to (26) has non-overlapping binary-valued allocation for each cell, i.e., αkj(f)αqj(f)=0, qj≠kj and αk
max Σj Σkj wkj(∫ {tilde over (c)}kj(f)αkj(f)df)
s.t. Σkj αkj(f)+Σkα{circumflex over (k)}ij(f)≦1, ∀ f,j;δkj≦∫ αkj(f)df≦γkj, ∀kj,j
α{circumflex over (k)}ij(f)=αkj; αkj(f)≧0, ∀f,kj,j. (27)
The objective function in (26) when evaluated for any optimal solution to (27) will yield a value that is within εΣj Σkj wkjγkj of the optimal value of (26). The problem in (27) and let denotes the partition obtained by taking the union of all partitions corresponding to all {{tilde over (c)}kj(f)}.
is a finite partition and in each frequency band in it, all {{tilde over (c)}kj(f)} are constants. Thus, there must exist a non-overlapping binary valued feasible allocation that is also optimal for (27). In sum, there must exist a non-overlapping binary valued feasible allocation that yields a value no less than εΣj Σkj wkjγkj from the optimal value of (26). Finally since ε can be made arbitrarily small the above discussion is true.
The problem of multi-cell subcarrier allocation under the equal power and bandwidth constraints (multi-cell type-2 resource allocation) can be formulated as follows,
max Σj Σkj wkj(Σn
s.t. Σkj Σn
Σnj; αkjln, βkjn ε{0,1}. (28)
Distributed Method for Resource Allocation
Next, low-complexity distributed processes which yield near-optimal solutions to the multi-cell resource allocation problems will be discussed. The first method is a greedy process, which consists of two steps. In the first step, the auction method is independently performed for resource allocation in every cell ignoring the constraint that each edge user and its corresponding virtual user(s) must share the same resource. The reward for allocating a subcarrier in a cell to any of its virtual users is set as zero in the first step, so no cell will allocate any resource to its virtual users. In the second step every edge user occupies the same subcarriers as its corresponding virtual user(s), by trading subcarriers in a greedy fashion. The detailed process is given below.
Process 4 Greedy Method with Auction Process for Mutli-Cell Type-1 Resource Allocation
Likewise, an auction-based greedy process can be used for multi-cell type-2 resource allocation problem. The main difference is that in the second step the sub-optimal Process-2 is run for each cell (with the alternate initialization if needed). Further, in the third step when a subcarrier is taken away from a user, its power is increased accordingly over each of its remaining assigned subcarriers.
In another subcarrier allocation scheme based on a combination of Lagrange dual method and the auction method, part of constraints is moved into the objective function to obtain a partial Lagrangian and the resulting problem of finding the dual function decouples into a several single-cell problems. To illustrate, in the multi-cell type-1 resource allocation, the dual function is obtained as:
max Σj Σkj wkj(Σl ckjlαkjl)+Σj ΣkΣl λ{circumflex over (k)}ijl(α{circumflex over (k)}kijl−αk
s.t. Σkj αkjl+Σk α{circumflex over (k)}ijl≦1, ∀l,j; nkjl≦Σl αkjl≦nkju, ∀kj,j; αkjl, α{circumflex over (k)}ijl ε{0,1}, ∀l,kj,j. (29)
The above optimization problem decomposes into subcarrier allocation problems for each cell. The subproblems can be exactly solved in an efficient manner by the auction process proposed in the previous section. After performing subcarrier allocations for each cell, let g(λ) denote the solution of the above problem for a given λ. These subproblems are coordinated by a high-level master problem, given by minλg(λ).
This dual master problem can be solved by a subgradient method in a distributed manner. Let s(λ): denote the subgradient of the function g(λ). The element of s(λ) with corresponding index {j,{circumflex over (k)}ij,l,ki} is given by {tilde over (α)}{tilde over (k)}
Process 5 Joint Action and Dual Decomposition for Multi-Cell Type-1 Resource Allocation
This dual method can be readily extended to solve the multi-cell type-2 resource allocation problem. The only difference is that Process 2 is run (with the alternate initialization if needed) to perform the type-2 resource allocation at the first step of each iteration. As in other dual decomposition methods for OFDMA, strong duality property is not guaranteed for finite N. Here, the third step described in the greedy Process 4 recovers a primal feasible solution upon convergence of the dual variables or after a predetermined number of iterations of the sub-gradient method. However, if at convergence the computed sub-gradient is a vector of all zeros, then the solution obtained in (29) is also a primal feasible solution and hence optimal. Also, the following proposition can help to computationally assess the performance of the proposed method after finite number of iterations.
After I iterations of the above combined auction and dual decomposition method, let Qp(I),Qd(I), and Q* denote the rates achieved by the obtained primal feasible solution, the dual solution, and the optimal solution, respectively, then Qp(I)≦Q*≦Qd(I).
In practice, if the gap between Qp(I) and Qd(I) is small enough, it is guaranteed that a near-optimal solution has been obtained. Furthermore, although the solution obtained by the joint auction and dual decomposition method is not necessarily optimal for finite N, for the multi-cell type-1 resource allocation problem, it is guaranteed to be asymptotically optimal in the limit as the number of subcarriers goes to infinity. Further, the solution derived by the joint auction and dual decomposition for the multi-cell type-1 resource allocation problem is asymptotically optimal in the limit as the number of subcarriers in every cell goes to infinity.
Next, the process selects the first element, say {n1, . . . , nK} in the queue, in which nk denotes the number of subcarriers allocated to user k. The process applies the procedure for type-1 resource allocation, with constant bandwidths {n1, . . . , nK} and constant power levels Pk/nk to perform the subcarrier assignment. The process sets ρ to be equal to the maximum of current best value and the obtained objective function value (214).
The process then selects the subcarrier with smallest reward, suppose it is the jth subcarrier and the associated user is kj (216). The process sequentially assigns this subcarrier to all other users except user kj, and performs equal power allocation and stores the obtained weighted sum rate. The process then locates among all elements in the above step satisfying the minimum and maximum bandwidth constraints, and adds those elements into the queue for which the corresponding weighted sum rates obtained in the previous step are larger than r (218).
The procedure is a greedy one in which at the initial step no sub-carrier is assigned to any user. Subsequently, at each step the process selects and assigns sub-carriers to a user (from the set of users which have not been selected yet) that results in the largest positive increase in the system utility. The process terminates when either all users have been selected or when no user (from the set of users which have not been selected yet) can offer an increase in the system utility,
Turning now to
Next, for each user, the process iterates in a loop (252). The process works on any user k not selected so far (254). For each user, the process sets the Modek to 0 and sets Maxk=0; n=1(256). The process then selects the top (or largest) n values out of {wkckin−v(i)} for all 1≦i≦N (258) and checks to see if anyone in the top n values is not positive (260). Next, the process sets I(k)n−1 (262) and then checks if Maxk is less than the sum of top n values (264). The process then sets the Modek to n and saves i(k,1), . . . ,i(k,Modek) which represent the subcarriers corresponding to the top n values. The process also set Maxk equal to the sum of top n values (266). From 264 or 266, the process increments n (268) and then checks if n is less than or equal to I(k) in 270). If not, the process proceeds to 272 to determine whether unselected users remain to be processed (272). Alternatively, if n is less than or equal to 1(k), the process loops back to 258.
From 272, the process determines q=arg max{Maxk} and checks if Maxq is greater than zero (274). Next, in 276, the process designates a user q as a selected user and assigns sub-carriers i(q,1), . . . ,i(q,Modeq) to user q after removing them from their previously assigned users if any. The process also sets v(i(q, j))=xqcq,i(q,j)Mode
The auction method for two resource allocation problems in uplink OFDMA networks. Both single-cell and multi-cell scenarios have been considered. The processes are designed for finite input alphabets and also account for non-ideal practical outer codes. In addition, power and bandwidth constraints motivated by practical concerns have also been handled well. The processes have very low complexity and can be implemented in a distributed fashion, when employed for multi-cell resource allocation. It is shown through both theoretical analysis and simulation studies that the system offers near-optimal performance. Several numerical examples have been done to demonstrate the performance of proposed Processs and substantiate the theoretical claims, as discussed in more detail in Provisional Application Ser. No. 61/035,400 filed on Mar. 11, 2008, the content of which is incorporated by reference. The result shows that Process 4 can improve the total system throughput considerably while Process 5 can further increase the system throughput but with an additional complexity. Both processes are distributed and of low complexity.
The above system applies an auction based approach to solve the single-cell resource allocation problem. The problem can be converted into a classical assignment problem and optimally solved by the auction method. Extended embodiments of the single-cell resource allocation procedure are applied to a multi-cell OFDMA uplink using an auction plus greedy method as well as a combined auction and dual decomposition approach. The system is low in complexity, as the complexity of the methods at each iteration increases only linearly with the number of subcarriers in the system. The method is much lower in complexity than the greedy method. The methods can obtain the optimal solution for the single-cell problem and can achieve solutions close to optimal for the multi-cell problem. Compared with the greedy method, the total single-cell throughput can be improved by up to 20%. The methods are well suited for parallel and distributed implementation. The resource allocation methods support practical modulation and coding scheme (MCS) rather than the commonly assumed Gaussian input alphabets and ideal outer codes. While Gaussian alphabets greatly simplify analysis, their main drawback is that the capacity with Gaussian input is unbounded in the signal-to-noise ratio (SNR) which results in too much power being allocated by a user to its stronger subcarriers.
The system may be implemented in hardware, firmware or software, or a combination of the three. Preferably the invention is implemented in a computer program executed on a programmable computer having a processor, a data storage system, volatile and non-volatile memory and/or storage elements, at least one input device and at least one output device.
Each computer program is tangibly stored in a machine-readable storage media or device (e.g., program memory or magnetic disk) readable by a general or special purpose programmable computer, for configuring and controlling operation of a computer when the storage media or device is read by the computer to perform the procedures described herein. The inventive system may also be considered to be embodied in a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.
The invention has been described herein in considerable detail in order to comply with the patent Statutes and to provide those skilled in the art with the information needed to apply the novel principles and to construct and use such specialized components as are required. However, it is to be understood that the invention can be carried out by specifically different equipment and devices, and that various modifications, both as to the equipment details and operating procedures, can be accomplished without departing from the scope of the invention itself.
This application claims priority to U.S. Provisional Application Ser. No. 61/035,400 filed on Mar. 11, 2008, the content of which is incorporated by reference.
Number | Name | Date | Kind |
---|---|---|---|
7586990 | Subramanian et al. | Sep 2009 | B2 |
20070226118 | Gill | Sep 2007 | A1 |
20080075032 | Balachandran et al. | Mar 2008 | A1 |
Number | Date | Country | |
---|---|---|---|
20090232074 A1 | Sep 2009 | US |
Number | Date | Country | |
---|---|---|---|
61035400 | Mar 2008 | US |