The present invention relates generally to the configuration of a cellular communication network and, more particularly, to a method of self-configuration of parameter settings for a cellular communication network.
In a mobile communication network, the coverage area of the network is divided into a plurality of cells. Conventionally, the configuration of cells and the assignment of certain parameters are set during cell planning and remain static once the system is deployed. As an example, frequencies may need to be assigned to the cells in a manner to reduce interference between cells. This requirement usually means that the same frequency is not reused in neighboring cells. As another example, the cells may be assigned locally unique reference signals (RSs), which are used by the mobile terminal to identify the cells and synchronize to the network.
The assumption that the network topology will remain static once a system is deployed is no longer true. It is now common for cells to be added or subtracted from a network after the system is deployed. For example, pico cells or other low power cells may be deployed in some areas of the network in order to improve coverage. Also, home base stations may be deployed in areas where network coverage is already provided. As new cells are added or subtracted, a different set of neighbor cell relationships is created.
Some mechanism is needed in order to prevent parameter conflicts between neighboring cells when new cells are deployed. For example, a new cell should not be assigned frequencies or RS patterns that conflict with existing assignments. Conventionally, the network administrator can manually assign parameters to the new cells in order to avoid conflicts. However, this is becoming increasingly more difficult as the number of cells in the network increases. In some cases, the network administrator must reconfigure existing cells in order to avoid conflicts. The cell planning that precedes the deployment of new cells can therefore be costly and time-consuming.
One way to simplify deployment of new cells is to make the network self-configuring in as many aspects as possible. U.S. patent application Ser. No. 13/072,496 filed Mar. 25, 2011 which is incorporated by reference herein in its entirety, discloses a general framework for updating system parameters when new cells are deployed. This patent application describes a distributed parameter update procedure based on the exchange of impact functions between neighboring cells. An impact function is a function describing how changes in a parameter setting of a first cell will affect a neighboring cell. When a parameter setting is updated by a cell, the cell exchanges impact functions with its neighboring cells. After the exchange of impact functions, the updating cell can compute a new parameter setting that minimizes the cumulative impact on its neighboring cells. A similar technique is used in other cells, so that as a whole the system gradually converges to an optimum or near optimum result.
The technique described above can be applied to a wide range of system parameters. In cases where the parameter setting is limited to a finite number of discrete values, the impact function can be easily tabulated and exchanged among cells. For a continuous parameter, such as transmit power, the same formulation can be used only if the parameter is quantized. If the quantization step is too large, quantization error may lead to performance degradation. If the quantization step is too small, the large amount of data to be exchanged among cells may exceed the capacity of the inter-cell connection or incur excessive delay.
Accordingly, there remains a need for a distributed parameter update procedure for updating parameters that can have continuous values.
Particular embodiments of the present invention provide methods and apparatus for updating, in a distributed manner, parameters of a network node that can have a continuous (non-discrete) value. When a parameter value is changed, a search is conducted of a parameter space to find a new parameter value that minimizes some cost function. The cost function is derived based on the current parameter settings in neighboring nodes. Certain embodiments of the present invention simplify the search process by localizing the search of the parameter space for a new parameter value to the vicinity of the current parameter setting. In some embodiments, the search is conducted along a line of steepest descent emanating from the current parameter setting.
Exemplary embodiments of the invention comprise a distributed parameter update method implemented by a network node in a mobile communication network. The method may be performed iteratively by the network node. In one exemplary method, a network node receives current parameter settings for each of one or more neighboring nodes in its local network. The current parameter settings may comprise settings computed in the previous iteration of the process. The network node then computes a gradient of a summed cost function for network node. The summed cost function is a function of the current parameter settings for the network node and the neighboring nodes in its local network. The network node then determines a revised parameter setting by searching the parameter space along the gradient of the summed cost function for a solution that minimizes the summed cost function. Once the new parameter value is determined, the network node sends its new parameter setting to the neighboring nodes in its local network. After exchanging new parameter settings with its neighbor nodes, the process may be repeated.
In another exemplary embodiment, the network node receives current parameter settings for each of one or more neighboring nodes in its local network. After receiving the current parameter settings, the network node determines a summed cost function as a function of the current parameter settings for network node and its neighboring nodes The network node then computes a gradient of the summed cost function for the network node. The network node then determines a revised parameter setting by searching the parameter space along the gradient of the summed cost function for a solution that minimizes the summed cost function. Once the new parameter value is determined, the network node sends its new parameter setting to the neighboring nodes in its local network. After exchanging new parameter settings with its neighbor nodes, the process may be repeated.
In another exemplary embodiment, the network node exchanges current parameter settings with the neighboring nodes in a local network of the network node. The network node then computes a partial derivative of a local cost function with respect to each of its neighboring nodes. A separate instance of the local cost function is computed for each neighboring node. Once the partial derivatives are computed, they are exchanged with the neighboring nodes. The network node then computes a gradient of a summed cost function at its current parameter setting based on the partial derivatives received from its neighboring nodes. The computed gradient is exchanged with the neighboring nodes. In the exchange, the network node receives gradients computed by each of its neighboring nodes. After this exchange, the network node computes a local impact function based on the gradients. A separate instance of the local impact function is computed for each of the neighboring nodes. The network node exchanges the local impacts functions with its neighboring nodes. After the exchange, the network node computes a summed impact function as a weighted sum of the impact functions of the network node and its neighboring nodes. The network node determines a new parameter setting based on the summed impact function and the gradient of the summed cost function. The new parameter value for the network node is then sent to the neighboring nodes. After exchanging new parameter settings with its neighbor nodes, the process may be repeated.
Other embodiments of the invention comprise a network node configured to implement the distributed parameter update procedures described above. In one embodiment, the network node comprises an interface circuit for communicating with one or more neighboring nodes in a local network of the network node, and a configuration processor for determining a parameter setting for the network node. The configuration processor is configured to implement the methods described as described above.
Certain embodiments of the present invention enable the use of a distributed parameter update procedure for parameters that can have a continuous value. The use of local impact functions in some embodiments requires the exchange of less information between network nodes as compared to a full description of the cost functions. Thus, bandwidth is conserved and latency is reduced. Further, the search of the parameter space is restricted to a line extending from the current parameter setting requires fewer processing resources.
Referring now to the drawings,
In the communication network 10, the nodes 20 may need to adapt a parameter, such as transmit power, responsive to short-term changes in the communication network 10. The parameter setting in each cell 12 will affect neighboring cells. For example, if a node 20 transmits at high power, it creates interference in the neighboring cells 12, which decreases capacity in the neighboring cells 12. On the other hand, if the node 20 transmits at a low power, it may be operating at less than maximum capacity. It should be noted that transmit power is a parameter that can have a continuous value between a minimum value (e.g., 0) and a maximum value.
A distributed parameter update procedure is used to coordinate the parameter settings between cells 12 in order to optimize some aspect of network performance. For example, the distributed parameter update procedure could be used to determine transmit power settings for the nodes 20 to maximize system capacity. The distributed parameter update procedure uses impact functions that describe how parameter settings in one cell 12 affects its neighboring cells 12. In the parameter update procedure, each node 20 computes a set of impact functions that describes how the parameter settings in neighboring cells 12 affects its own cell 12 and exchanges the impact functions with its neighboring cells 12. The node 20 in each cell 12 uses the impact functions received from nodes 20 in neighboring cells 12 to determine its own parameter setting.
The distributed update procedure 50 applies in a straightforward way to parameters that have a finite set of discrete values. However, application of the distributed parameter update procedure to parameters that can have a continuous value is more difficult. Examples of continuous value parameters include the transmit power of a node 20 and antenna weights for nodes 20 in a multiple-input, multiple output (MIMO) system with multiple antennas. One of the problems in applying the distributed parameter update procedure to such parameters is the exchange of cost functions. For continuous variables, the amount of information needed to fully describe the cost function makes the exchange of information costly in terms of bandwidth and latency. Another problem is that the parameter space is quite large. Performing a full search of the parameter space for an optimum solution would be time-consuming and would require significant processing resources.
In one or more embodiments of the present invention, a node 20 restricts its search of the parameter space to a one dimensional subspace in the vicinity of the current parameter setting. In one exemplary embodiment, a node 20 searches the parameter space along a line of steepest descent emanating from the current parameter setting. This approach reduces the amount of information required to describe the behavior of the cost function in the vicinity of the current parameter setting, and requires less information to describe than the full cost function. Therefore, it is possible to apply the distributed parameter update technique to variables that have continuous values.
Introducing the notation used in the remainder of this description, each cell 12 in the communication network 10 is identified by a unique index i, and a one-to-one correspondence between nodes 20 and cells 12 is assumed. At a given time instant n, a parameter vector pi associated with node i can be set by the node to adapt to changes in the surrounding environment. If the parameter is the transmit antenna weights in a MIMO system, it can be set to any complex number. If the parameter is the transmit power of a node 20, it can be set to any non-negative number. By convention, variables are denoted by bold face type and the values it takes by normal lower case letters, such as pin, where the superscript n indicates the time index.
Each node i has a list of neighboring nodes determined by a proximity measure such as distance or intensity of radio interference. The determination of the neighboring nodes is not a material aspect of the invention, and techniques for determining neighbor sets are well-known in the art. In general, the neighbor set is determined based on received signal measurements. Nodes whose signal strength measurements exceed a threshold are included in the neighbor set. The set of neighboring nodes of node i is denoted i. It is assumed that the neighbor relationship is reciprocal, i.e., jε
i if and only if iε
j. The union of the neighbor set
i, with node i is referred to as the local network with respect to node i and is denoted
i.
Each node if can communicate with its neighboring nodes in i to exchange parametric information required to evaluate a cost function defined for its local network
i. The cost function is given by
(ipj|ε
i)=
(pi,pj|jε
i) Eq. (1)
After exchanging information with its neighboring nodes, each node i computes a new parameter value that minimizes a summed cost function given by:
That is, each node i searches the parameter space and selects its parameter value as the one that minimizes the weighted sum of the cost functions corresponding to the local networks of the nodes in i (the local network of the central node i), assuming all other parameters remain the same as previously communicated at time instant n. The updated parameter value is then communicated to the neighboring nodes while at the same time the updated parameter values of neighboring nodes are also received.
For continuous parameters, such as transmit power, a search over the entire parameter space for the optimum solution of pi would be computationally complex and time-consuming. Instead of searching over the entire parameter space for a globally optimum solution of pi as described above, a simplified search may be implemented wherein the search of the parameter space is made in each iteration along the direction of the steepest descent. This approach restricts the optimization problem to a local one that is easier to define and thus more efficiently communicated between neighboring nodes.
i (block 110). The current parameter settings are those computed in the previous iteration. Network node i then computes a gradient of a summed cost function Cin(pi) for network node i (block 120). The summed cost function Cin(pi) is a function of the current parameter settings for network node i and the neighboring nodes
i in its local network
i. Network node i then determines a revised parameter setting by searching the parameter space along the gradient of the summed cost function for the solution that minimizes the summed cost function (block 130). Once the new parameter value is determined, the network node i sends its new parameter setting to the neighboring nodes
i in its local network
i (block 140). After exchanging new parameter settings with its neighbor nodes
i, the process is repeated.
In a first exemplary embodiment, the summed cost function given in Eq. (2) is rewritten as:
The dependence on the neighbor nodes' parameter is dropped to highlight the fact that the cost function is a function of its own parameter variable pi. By convention, the parameter pi is a column vector if it consists of more than one variable. The gradient of the cost function in Eq. (2) is a row vector given by:
The elements of Eq. (4) are partial derivatives of the summed cost function with respect to the corresponding variables in the parameter vector. The gradient evaluated at pin is
where ( )T is the vector (or matrix) transpose. Therefore, the gradient γin is a column vector with the same dimension as the parameter pi.
In this embodiment, it is presumed that the cost function is known and that the gradient of the cost function can be evaluated at the current parameter value. The distributed parameter update procedure updates the parameter value by searching for a better solution than the current one along the direction of a line segment defined by pin−γinα, where α is a nonnegative scalar. The scalar a that may be upper bounded by the constraints of the problem. In other words, the network node i seeks the a value that minimizes the cost function:
The network node i then computes the new parameter value according to:
pin+1=pin−γin{tilde over (α)} Eq. (7)
The updated parameter values are then exchanged among neighboring nodes before the next iteration cycle starts.
i are known or can be determined. Network node i receives current parameter settings for each of one or more neighboring nodes
i in its local network
i (block 210). After receiving the current parameter settings, network node i determines a summed cost function according to Eq. (3) given the current parameter settings (block 220). Network node i then computes a gradient of the summed cost function Cin(pi) for network node i (block 230). The gradient may be computed according to Eq. (5). Network node i then determines a revised parameter setting by searching the parameter space along the gradient of the summed cost function for the solution that minimizes the summed cost function (block 240). Once the new parameter value is determined, the network node i sends its new parameter setting to the neighboring nodes
i in its local network
i (block 250).
In many cases, each parameter pi is also constrained to be in a feasible set Si. In this case, a function projecting the perturbed parameter value onto Si may be used. In cases where the parameter value is limited to a predefined set, the parameter is given by:
pin+1=QS
where QS
The evaluation of the summed cost function given in Eq. (3) requires knowledge of the individual cost functions A
i. If this information is not available at node i, it needs to be communicated from the neighboring nodes
i. This exchange may be too costly in terms of bandwidth or latency if the cost functions are very complicated, or if the cost functions do not have a functional form (i.e., cannot be represented by an Eq.). Moreover, the cost function may change as network topology or propagation environment evolves over time, making the exchange of local cost functions even more inefficient.
Fortunately, the steepest descent approach requires only knowledge of the cost function along the line of gradient in the neighborhood of the current parameter values. This restricted function of a single variable, referred to herein as the “local cost function,” is easier to describe than the functional form of the cost function. Therefore, in a second embodiment of the invention, each network node i exchanges information describing its local cost function in the vicinity of its current parameter setting with its neighboring nodes.
In the second embodiment, each node i first computes the partial derivative gj→in+1 of its local cost function for each of its neighboring nodes. The partial derivative may be computed according to:
After exchanging the partial derivatives, each node i will have sufficient information to compute its search direction. The search direction can be computed according to:
Once the search direction is determined, it is then communicated to the neighboring nodes i. That is, each node i exchanges the computed search direction computed in Eq. (12) with each of its neighboring nodes
i. The exchanged search direction information is used at each node i to determine a local impact function for each of its neighboring nodes
i. In one embodiment, the local impact function is computed according to:
Ij→in+1(α)=(pjn−γjn+1α|pk=pkn,∀kε
i−{j}) Eq. (13)
This local impact function describes the contribution of the parameter set by each neighboring node i to the summed cost function
of node i. Since it is a scalar function of a single variable α, it can be more concisely represented than the local cost function itself. A separate instance of the local impact function is computed for each neighboring node
i.
After exchanging the local impact functions with it neighboring nodes i, the network node i can compute a summed impact function describing how changes to its own parameter setting will impact its neighboring nodes
i. The network node i seeks the α value that minimizes the following summed impact function:
The network node i then computes the new parameter value according to:
pin+1=pin−γin+1{tilde over (α)}. Eq. (15)
The updated parameter values are then exchanged among neighboring nodes and the next iteration of the distributed parameter update and the process is repeated using the new parameter settings.
i are not known. Network node i exchanges current parameter settings with the neighboring nodes
i in a local network
i (block 305). The network node i then computes a partial derivative of a local cost function
with respect to each of its neighboring nodes (block 310). The partial derivative may be computed according to Eq. (11). A separate instance of the local cost function
is computed for each neighboring node
i. Once the partial derivatives are computed, they are exchanged with the neighboring nodes
i (block 315). The network node i then computes a gradient of a summed cost function at its current parameter setting based on the partial derivatives received from its neighboring nodes
i (block 320). The gradient may be computed according to Eq. (12). The computed gradient is exchanged with the neighboring nodes (block 325). In the exchange, node i receives gradients computed by each of its neighboring nodes. After this exchange, the network node i computes a local impact function according to Eq. (13) based on the gradients (block 330). A separate instance of the local impact function is computed for each of the neighboring nodes. The network node i exchanges the local impacts functions with its neighboring nodes (block 335). After the exchange, the network node i possesses a local impact function describing how changes in its parameter setting will affect each of its neighboring nodes. The weighted sum of these impact functions yields a summed impact function of the variable α (block 340). The network node i determines a new parameter setting based on the summed impact function and the gradient of the summed cost function (block 345). More particularly, the network node searches for the value of α that minimizes the summed impact function and computes the new parameter value according to Eq. (14) and Eq. (15). The new parameter value for the network node is then sent to the neighboring nodes (block 350).
In the embodiment shown in
A practical example of a particular embodiment of the present invention will now be described to illustrate one possible use of the invention. Power control is a commonly encountered problem in a cellular network. If a node is transmitting at very high power, it may create too much interference to neighboring cells. If its transmit power is set too low, it may not be operating at its maximal potential. Ideally, a central node can be employed to coordinate the transmit power from the multiple cells so that the overall network-wise capacity is maximized while maintaining a certain minimal quality of service for the individual cells. In practice, however, such a central node may not be available and a distributed solution is needed.
i centered on node P0, which for purposes of this example is assumed to be a base station. The neighbor set
0 for node P0 includes nodes P1, P2, P3, P4, P5, and P6. A user terminal being served by node P0 receives its desired signal with a power level of p0g00, where p0 is the transmit power of node P0 and g00 is the path gain between node P0 and the user terminal. In addition, it also receives interference from at least the six cells 12 in its neighbor list at a level of pjg0j, where pj is the transmit power of neighboring node j and g0j is the path gain between neighboring node j and the user terminal. Based on the information available to node P0, the capacity of the user terminal it serves is given by:
where N0 is the noise variance.
The goal of the distributed parameter update method is to maximize the summed capacity of all the cells 12 in the network 10 subject to a power constraint at each node and a certain minimal requirement constraint for the individual cells 12. Such constraint can be, for example, a limit on the highest interference each user terminal experiences. The interference may be expressed by the ratio:
The requirement is, therefore, to keep the ration given by Eq. (16) below a predetermined threshold λ for all nodes jε0.
The local cost function defined in Eq. (1) is chosen to be the negative of the capacity and is given by:
The local cost functions for any jε0,
A
The partial derivatives of Eq. (18) can be easily derived, and are given by:
for kε0, k≠0. With the local cost function defined and its partial derivative derived, the distributed parameter update procedures of
In some embodiments of the invention, the constraint given by Eq. (17) on highest interference level, the summed cost function over the range of valid α in the line search given in Eq. (6) and Eq. (14) can be closely approximated by a linear function, thereby further reducing the amount of information exchanged among neighboring nodes.
The disclosed embodiments of the present invention enable the use of a distributed parameter update procedure for parameters that can have a continuous value. The use of local impact functions in some embodiments requires the exchange of less information between network nodes as compared to a full description of the cost functions. Thus, bandwidth is conserved and latency is reduced. Further, the search of the parameter space is restricted to a line extending from the current parameter setting, requires fewer processing resources.
The present invention may be carried out in other ways than those specifically set forth herein without departing from essential characteristics of the invention. The disclosed embodiments are thus to be considered in all respects as illustrative and not restrictive, and all changes coming within the meaning and equivalency range of the appended claims are intended to be embraced therein.
Number | Name | Date | Kind |
---|---|---|---|
20040266442 | Flanagan et al. | Dec 2004 | A1 |
20090073893 | Vu Duong et al. | Mar 2009 | A1 |
20100034157 | Stolyar et al. | Feb 2010 | A1 |
20110176433 | Monogioudis | Jul 2011 | A1 |
Number | Date | Country | |
---|---|---|---|
20130109421 A1 | May 2013 | US |