JP2012522446A - Tree search method by depth-first search for detecting MIMO received signal - Google Patents

Abstract

  A tree search method based on depth-first search for detecting a received signal of a multi-input / multi-output channel based on a tree search, in which a leaf node of the search tree represents a possible transmission bit and is transmitted with the highest likelihood. Determination of the bit considered to have been performed and the reliability value of that bit, which is determined based on the state information of the leaf nodes detected during the tree search, and the search space for the tree search is We propose a limited method based on the state information of leaf nodes detected during the search. In the present invention, the search space for the tree search is limited independently of the processing of the state information for calculating the reliability value, and more specifically, the leaf node state information detected during the tree search Are performed using n columns of tapels, and the maximum value of the state information of these tapels is determined as the radius of the search space. The proposed method makes it possible to significantly reduce the number of operations and the number and scale of hardware units necessary for limiting the search space, which improves the detection performance and area efficiency, and It can be adopted for mobile systems.

Description

  The present invention is a tree search method based on a depth-first search for detecting a received signal of a multi-input / multi-output (MIMO) channel based on a tree search, wherein a leaf node of the search tree represents a possible transmission bit. , Including the determination of the most likely transmitted bit and its reliability value, which is determined based on leaf node state information detected during the tree search, The present invention relates to a method in which a search space for search is limited based on state information of leaf nodes detected during a tree search.

  The ever-increasing data rate in mobile radio calls for improved usage of available bandwidth. Therefore, in order to improve spectral efficiency, a system (multi-input multi-output (MIMO) system) having a plurality of transmission antennas and reception antennas is increasingly used. As is generally known, the complexity of an optimal MIMO receiver (APP (posterior probability) detector) increases exponentially with the number of bits per transmitted symbol. The tree search algorithm is a very attractive technique for achieving near optimal performance to the extent that detection complexity is acceptable for multiple antenna systems.

  In recent years, such algorithms have been energetically studied and initially assumed to be implemented in hardware based on application specific integrated circuits (ASICs). In the mobile radio generation to be further studied in the future, in addition to MIMO detection, iterative detection algorithms, interference suppression, variable system parameters, network MIMO and new coding methods (eg CLDPC / Convolutional Low Density Parity Check Codes) will be used. . They greatly increase the architectural requirements for implementing the tree search method in terms of computing power, power consumption and reconfigurable form. However, the developed solutions are far from the solutions that satisfy the increasing requirements of such next generation mobile radio systems.

  Non-Patent Document 1 describes the principle of an M algorithm or a K-Best tree search algorithm based on a breadth-first search having an extension for calculating a reliability value (so-called soft output value) and a method of applying it to an ASIC. Has been. In an implementation for a 16-QAM (Quadrature Amplitude Modulation) system (4 × 4 MIMO) with four transmit antennas and four receive antennas, the hardware for each of the eight tree planes processed in the order of the actual alternative system. Components exist and are selected such that they can be processed in parallel with an expansion unit for computing soft output values by pipeline processing. By doing so, the detector can receive at data rates up to 106 Mbit / s, but the complexity has increased significantly with the number of computational units and has been greatly simplified. Deterioration of accuracy occurs due to the soft output value calculation.

  As described in Non-Patent Document 2, the latest detector based on ASIC for depth-first search using the sphere detection algorithm is also about 10 Mbit with almost optimal detection quality (accurate calculation of soft information). A throughput of / sec is achieved. Thus, in addition to further optimization of such methods, multiple parallel detectors may be employed to achieve data rates that clearly exceed the 100 Mbit / s pursued by mobile radio. is necessary.

FIG. 1 illustrates a system model of a MIMO system having _NT transmit antennas and N _R receive antennas. This system model is represented as a complex value. Vectors are highlighted with small symbols shown in bold, and matrices are highlighted with large symbols shown in bold. Vector or matrix transformation is indicated by the superscript T, and Hermitian transformation (conjugate complex transformation) of the vector or matrix is indicated by the superscript H.

Represents a set of complex numbers. Real number operation is

And the complex operation is

It is shown in

Represents the Euclidean norm of each argument.

  FIG. 1 shows the main elements of the transmission section. The description here is for the baseband, that is, the discrete state. Data is transmitted in blocks, and the impulse response filters of the transmitter and receiver are not considered separately in this embodiment, but can be integrated into the channel model. On the transmitting side, a vector u in which the data bits are independently distributed in the same form is encoded (114) with the outer channel code, and the flow obtained from the vector c ′ is bit-interleaved (116), and the block It is divided into c. The corresponding bits are mapped (118) to complex symbol constellations for transmission.

Represents a complex system matrix of size (N _T × N _R ), which includes transmission channels and transmission and reception side filters (eg, channel adaptation filters, RAKE / pre-RAKE filters, inverses) And is known in a receiver for detection. The data block c = (c (1),..., C (N _T )) ^T is possibly composed of N _T symbols, each consisting of L bits, encoded and interleaved before transmission. Is done. For modulation, the element is

The mapping ζ _m = map (c _m ) of the m-th bit vector _cm to the corresponding symbol ζ _m can be performed by, for example, gray labeling. Q = 2 ^L represents the order of the symbol, and typical values of Q are 4, 16, 64, 256. The in-phase or quadrature component of the symbol is indicated by x _I and x _Q. Transmit signal in a set of possible transmit symbols

The received signal by multiplication with the system matrix

And the received signal is a complex-valued additional noise whose noise power density is N ₀

Are superimposed, and y = Hx + n.

  FIG. 2 illustrates an example of symbols and bit values based on 64-QAM symbol arrangement by gray labeling. FIG. 3 illustrates a 16-QAM example of a received symbol arrangement that is distorted by transmission and serves as a detection base. Based on the received symbol y, the transmitted data is estimated. Specifically, the estimation is performed by using the detector 122 shown in FIG. 1 and the decoding unit 128 corresponding to the preprocessing unit on the transmission side, and detection / decoding is performed. The process can also be performed iteratively (126).

  Thereafter, the decoded data is input to the binary data receiver 132 via the hard decision block 130 by a well-known method.

  The role of the detector 122 considered here, in which the present invention can be implemented, is that the bit c that appears to have been transmitted with the highest likelihood and the logarithmic ratio of those bits with one bit being “1”. ”Or“ 0 ”, or“ +1 ”or“ −1 ”, and reliability information (L value) associated with it. They are calculated from the received symbols, channel estimates, calculated noise power and possibly available initial information. The Bayes' theorem, the statistical independence of transmitted bits, the multi-dimensional Gaussian noise density function, and the Max-log approximation can represent the detection process as follows.

In the Max-log approximation, a surface-specific distance matrix is established, and H = QR, y ′ = Q ^H y holds. For each detected antenna (i = 1... N _T ), there are a plurality of possible transmission symbols, and therefore a plurality of transmission bits for individual bits of the transmission vector determined by the triangular structure of the analyzed system matrix. The pursuit of hypotheses or alternative hypotheses can be interpreted as a search based on tree-structured leaf nodes (i = 1). Various tree search methods are illustrated in FIG.

In this case, the search tree has a plurality of planes i, and each plane represents one estimated transmission symbol. In each of the tree planes (i = 1... N _R ), there are Q possibilities corresponding to the respective selected symbol arrangement with respect to x _i . Such a possibility is represented by a tree node, using the interference occurring between the corresponding symbol and the remaining symbols (1, _... , I−1) that have not yet been estimated. It is guided to the child node of (i-1). Therefore, the leaves of the tree represent the estimated complete transmission data block.

The estimation of the transmitted symbol x using QR analysis of the system matrix can be performed, for example, by linear zero forcing (ZF, complete interference suppression) or linear least mean square error (MMSE, least square) in the extended channel matrix. Equivalent to the use of an error criterion) and can be used for real-valued alternative models, sorted QR analysis (SQRD) or lattice reduction system matrices as well. At the time of estimation, the transmission point shifts within the vicinity of the reception point in the I / Q plane due to distortion (see FIG. 3), and searches for a distorted transmission signal with the smallest deviation across all antennas. It is performed by repeating the inverse permutation of transmission symbols already estimated and the calculation of Λ (c, y, L _a (c)) by the cumulative distance metric λ _i belonging to it. Because of the many symbols that can be assumed, it is necessary to use special search methods to limit the computational complexity. This is done by limiting the search path in the search tree. In this case, as described in Non-Patent Document 3, the tree search method can be roughly classified into three types of algorithms: depth priority search, metric control search, and width priority search.

In the depth-first search (see FIG. 4B) of the sphere detection (SD) algorithm, as described in Non-Patent Documents 4 to 6, the search space is as much as possible to the hypersphere (sphere) around the received symbol. We try to limit it quickly and thereby reduce the number of calculations performed. In this case, the radius R of the sphere adapted by the calculated leaf node distance metric λ ₁ functions as a parameter for such restriction. After initial consideration of the unrestricted search space (initialization with R = ∞), for the so-called baby point (see FIG. 4A), for example, R: = Limit search space to λ _{1, Babai} . The tree planes are then considered in reverse order, and a depth-first search is performed on each unexpanded tree node in the current search sphere (λ ₁ ≦ R). In this case, the selection of the nodes to be considered is advantageously the so-called Schnorr-Euchner enumeration method, which usually always considers the next prospective node as the next node, as described in Non-Patent Document 7. Is done by. The search and the search space adjustment by determining a new leaf node is performed until all nodes in the sphere are considered.

In contrast to the depth-first search, in the search based on the matrix, as described in Non-Patent Documents 8 and 9, basically, the number of metric calculations is immediately minimized without taking a detour due to the limitation of the search space. Yes. In this case, we list the metric λ _i of all considered nodes in a list, sort them, and extend to the node with the smallest metric. The first leaf selected for expansion (i = 1) corresponds to the ML / MAP solution. The search performed further provides alternative assumptions belonging to it.

  The third type of tree search method includes so-called breadth-first search such as M algorithm and K-Best algorithm as described in Non-Patent Documents 10 and 11, for example. In this case, in each face i + 1 of the tree, for the selected node, calculate the next promising node (face i) and select the M or K-Best solution for the next face calculation. . The basic flow is illustrated in FIG. 4 (c) for M or K = 3.

A special variation of the tree search method by depth-first search for detecting a MIMO received signal is, for example, a so-called list sphere detection (LSD) algorithm as described in Non-Patent Document 6. The conventional processing flow of such an algorithm is illustrated in the data flow diagram in FIG. This algorithm is executed as follows after initialization in step 502.
1. If the search has not yet been completed (decision 540), i.e., if the search tree has not been completely searched, or if a stop condition such as the maximum number of clocks has not been met, then continue this algorithm, or If there is not, the following 6. The process is executed.
2. The next step is to further consider the in-plane child nodes from time to time,
2.1 When considering a new surface (or antenna) of a tree (decision 503), the interference of already estimated symbols must be removed from the received symbol (step 504), and the distance to the child node must be calculated. (Step 505). In this case, for selection of the most probable node, for example, calculation of all child nodes and sorting based on the likelihood (step 506) are necessary.

2.2 If we have already considered and sorted the child nodes of the current tree face (and therefore the same presumed parent node), this process requires no further processing and is processed next A parent node is selected (step 508).
3. Next, in some cases, the search space can be limited,
3.1 When the tree leaves are reached (decision 510), the search sphere can be further limited in some cases. The approach for this is to store the signal candidates important for the search sphere and to calculate the search sphere by sorting the signal candidates (step 512) associated with the calculation of the preferred sphere radius (step 514).

3.2 If the leaves of the tree are not considered in the current plane (i ≠ 1), in this variation, the flow is shifted to the determination question 520 without further limiting the search sphere.
4). Next, decide the surface of the tree to consider,
4.1 If the leaf of the tree is reached, the surface i of the tree is raised by 1 (step 530).

4.2 If the tree leaves are not considered in the current plane (i ≠ 1), if all the nodes of the current search plane existing in the search sphere have already been considered, the tree plane is raised ( “Yes” in step 520). Otherwise, select the most probable and unconsidered node (step 522) and lower the face of the tree for that node (step 524).
5. Above 1. Continue to search for trees.
6). In order to finish the tree search, reliability information is calculated from the detected leaf nodes and stored (step 542).

  Even in this depth-first search algorithm and other methods, it is important to calculate assumptions and alternative assumptions and their distance metrics in order to calculate reliability information. In that case, with respect to the tree search method to be used, it is found that promising leaf nodes are found for both individual bit assumptions and alternative assumptions, and that they can be used to calculate reliability information. Must be guaranteed. At the same time, the search should have minimal complexity. For this purpose, the number of operations to be executed and the number of tree nodes to be considered are important. Therefore, advantageously limiting the search space is important in reducing computational complexity.

  In order to limit the search space but still calculate sufficiently promising alternative assumptions, there are basically two known methods in depth-first search besides explicit search methods for alternative assumptions. .

In the first method, the limitation of the search space is calculated from the radius (R = λ _Hypothese ) calculated from the hypothesized distance metric or the list sphere detection (LSD) algorithm described in Non-Patent Document 6. This is done with a radius (R = λ _{L-ter — Kandidat} ) calculated from the most promising signal candidates. In this case, the leaves of the tree are included in the signal candidate list along with their distance metric. The leaves included in the list are used for calculation of reliability values after searching. For a limited list of size L, to ensure that the most promising signal candidates are always stored in the list, the newly determined leaves are combined with the signal candidates stored in the list. Sorted in ascending order of distance metrics, only the L leaves with the smaller distance metrics are left and saved in the list. Following such a search, reliability information is calculated from the signal candidate list. As long as a sufficiently promising signal candidate can be calculated as the search progresses, the number of stored signal candidates limited by the size of the list is critical for the quality of the method. A tree search method that detects a list that is too small or a small number of promising leaves degrades the quality of the reliability information. In this case, it is disadvantageous that the complexity increases with the size of the list, which is related to the required hardware units as well as the computational complexity required to process the list. Each list element requires multiple storage elements. Furthermore, as the search progresses, the list must be sorted each time a new leaf is detected, thereby dramatically increasing hardware complexity. Furthermore, when the worst signal candidate in the list is used to determine the search space, the search complexity increases with the size of the list, and independent adjustment of search accuracy and search complexity is not possible. It becomes possible.

  A second method for limiting the search space is to introduce a radius for each bit as described in Non-Patent Document 12. In this case, the goal is to find the most promising tree leaf for every possible bit value (0/1). Thus, the search space depends on the bit value of the node being considered at any given time, and is expanded to a node that can always derive new or alternative assumptions as follows (ie, the child nodes of that node) Search for).

Where λ _{Gegenhyp., Cj} is the most promising alternative hypothesis distance metric of bit c _j , c _i is the bit value of the current node, and c _j ^h is the hypothesis of bit c _j Bit value. Each time a leaf is found, a distance metric is saved for the bit represented by that leaf. When one distance metric already exists for one bit, the smaller one is always left. As a result, this tree search is always extended to a node that can derive a smaller distance metric for at least one of the bits corresponding to the search path. Therefore, the search radius is different for each bit. This ensures that the most promising assumptions and the most promising alternative assumptions are always found when searching. Furthermore, the number of intermediate states to be saved for this purpose is given fixedly.

This method certainly requires a small memory capacity, but fundamentally changes the search flow. However, by finding the most promising leaves for each bit and extending the search space with a radius that is R (c _i ) >> λ _Hypothese , the number of tree nodes to be analyzed in the search space is also dramatically increased at the same time. To increase. The same is true for the complexity of determining the search space from time to time. In order to alleviate such effects, radius restrictions, so-called clipping, are often introduced, and a maximum value is set for the search space. By doing so, the tree search method finds the most promising tree leaves in a given search space, but in a sufficiently large search space, it still has to analyze a relatively large number of nodes, so this method Can only be used in a limited way for efficient calculation.

  The goal of this constant effort in the field is to reduce the complexity of the algorithm and hence the complexity of the tree search application to the hardware for signal detection without significantly degrading the accuracy of the search. is there.

"Verfahren zur baumsuchbasierten Detektion von Empfangssignalen" "Verfahren und Anordnung zur Auswahl von Ebenen bei Baumsuchalgorithmen der Tiefensuche"

G. Nilsson, P. Nilsson: "Algorithm and Implementation of the K-Best Sphere Decoding for MIMO Detection", IEEE Journal on Selected Areas in Communications, Vol. 24, Nr. 3, P.491-503, March 2006 C. Studer, A. Burg, H. Boelcskei: "Soft-output sphere decoding: Algorithms and VLSI implementation", eingereicht fuer IEEE Journal on Selected Areas in Communications, April 2007 J. Anderson & S. Mohan in "Sequential Coding Algorithms: A Survery and Cost Analysis", IEEE Transactions on Communications, Vol. 32, Nr. 2, P.169-176, February 1984 R. Fano in "A heuristic discussion of probabilistic decoding", IEEE Transactions on Information Theory, Vol. 9, P.64-74, April 1963 M. Pohst in "On the computation of lattice vectors of minimal length, successive minima and reduced bases with applications", ACM SIGSAM Bulletin, Vol.15, P.37-44, 1981 B.M.Hochwald and S.ten Brink in "Achieving near-capacity on a multiple-antena channel", IEEE Transactions on Commuinications, Vol.51, P.389-399, March 2003 C.P. Schnorr and M. Euchner in "Lattice basis reduction: Improving practical lattice basisreduction and solving subset sum problems", Mathematical Programming, Vol.66, P.181-199, August 1997 F. Jelinek in "Fast Sequential Decoding Algorithm Using a Stack", IBM Journal of Research, Vol.13, P.675-685, 1969 J. Hagenauer and C. Kuhn in "The List-Sequential (LISS) Algorithm and its Application", IEEE Transactions on Communications, Vol.55, P.918-928, May 2007 J. Anderson and S. Mohan in "Sequential Coding Algorithms: A Survey and Cost Analysis", IEEE Transactions on Communications, Vol.32, Nr.2, P.169-176, February 1984 S. Haykin, M. Sellathurai, Y. de Jong and T. Willink in "Turbo-MIMO for wireless communications", IEEE Communications Magazine, Vol.42, P.48-53, October 2004 J. Jalden and B. Ottersten in "Parallel Implementation of a Soft Output Sphere Decoder", Asilomar Conference on Signals, Systems, and Computers, November 2005 E. Zimmermann and G. Fettweis in "Unbiased MMSE Tree Search Detection for Multiple Antenna Systems", International Symposium on Wireless Personal Multimedia Communications (WPMC'06), September 2006

  It is an object of the present invention to reduce the number of tree nodes to be considered and hence the complexity of the calculation for calculating the distance metric and the application to hardware with comparable or improved search accuracy. It is to modify the tree search algorithm. In this case, it is assumed that this method has a simple structure and application form, and can realize high performance as well as high flexibility of the search method.

  The solution according to the invention for this problem is described in claim 1. Advantageous refinements are described in the dependent claims and are described in detail below in connection with the invention.

  The present invention presents a method with reduced complexity in a tree search method by depth-first search for calculating a received signal of a multi-input / multi-output (MIMO) channel based on a tree search. This method is characterized in that the search space for the tree search is limited independently of the processing of the state information for calculating the reliability value. The search space is limited by n columns of tuples in the leaf node state information detected during the tree search, and the maximum value of the state information of these tuples is determined as the radius of the search space. . By separating the search space limitation reliability value calculation according to the present invention, the size and complexity of the node analysis, the quality of the detected alternative assumptions, the cost of limiting the search space, the cost of calculating reliability information, etc. Can greatly reduce the problem of tree search. In this case, the state information stored in the n columns of tapels can be the distance metric of the leaf node of the search tree.

  In the present invention, during a tree search, leaf nodes that are found within the current search radius and have more promising state information than the least promising value of the n-column tapels are captured in the n-column tapels. Thus, the least promising element of the n-row tapel can be replaced. Doing so ensures that those tuples always contain the leaves of the n trees with the smaller distance metrics found so far.

  In an advantageous implementation of the invention, the n columns of tapels can be generated as a sorted list of list elements 1 to n, which are arranged according to the value of the state information. The list elements can be arranged in ascending order of values, for example, with the first position of the subset being the smallest value and the nth position of the subset being the largest value, depending on the distance metric.

  In an advantageous implementation of the invention, an n-column tapel can be implemented as a subset of a list of magnitudes where L> n, and the list of magnitudes L as a whole is for calculating confidence values. The signal candidate list can be represented together with the state information. The list of size L has been sorted to at least element n + 1, and the elements of the new possible tuple are incorporated in the n + 1th position of the sorted subset.

  Next, to limit the search space, the nth list element of the list of size L can be used, and the position of n (1 <n ≦ L) in the list of size L is variable. . The location of the radius at the nth position of the list of assumptions and alternative assumptions sorted by ascending order of the distance metric ensures that the better n leaves of the search tree are detected. Thereby, the quality of the alternative assumptions is guaranteed and the list can be selected small accordingly. Furthermore, it limits hardware complexity and minimizes the number and scale of operations required to sort the list. Replacement of a single tuple with a variable can be done by changing the size of the subset of storage elements defined by the use of available resources at a variable rate and the number to sort. By doing so, it becomes possible to adapt the complexity of the search and the accuracy of the search to the ambient conditions before or during the execution time of the tree search. Increasing tuples will detect more promising leaf nodes. In this case, a large number of predetermined promising leaf nodes increase the likelihood of calculating more promising state information for calculating reliability information, directly or indirectly, depending on the search method used, Therefore, on average, the search accuracy is improved.

  In an advantageous implementation, the state information used to calculate the confidence value is stored separately from the n-column taples used to define the search space. Particularly advantageously, storing state information separately for calculating reliability information is done bit by bit in a separate list. In this case, the state information is stored for each bit separately for each possible bit value and for all bits, or in an alternative implementation, for each transmitted symbol, It can be done in common for all bits and separately for the corresponding bits of alternative assumptions. By storing each bit, all leaves determined according to the progress of the tree search can be incorporated into the determination of the likelihood value. By separately processing the state information for calculating the reliability information, the processing can be optimally performed without affecting the progress of the search. Compared to the traditional list-based method where some alternative assumptions are discarded by being locked out by restricting the size of the list during the tree search, the present invention separates the distance metric for each bit separately. Because of the preservation, it can always be assured that the smaller distance metric found for one bit can also be used for the calculation of alternative assumptions. In this case, clipping of the value for the tree search or after the search is not affected. By incorporating all detected leaf nodes when storing the distance metric for each bit, loss of related state information during the progress of the tree search can be prevented, and the accuracy of the search can be improved.

  Therefore, compared to the search by tuple and the calculation of reliability information from the limited signal candidate list, it is possible to reduce the tuple as intended to limit the search space, when the detection performance is the same. Is possible. As a result, the required memory capacity and the computation required for sorting the list are reduced. The use of a radius for each bit eliminates the need for a costly search space. The method allows computing reliability information in parallel from a given list by using a bit-by-bit distance metric. Therefore, a search based on the most promising alternative assumption in other list-based methods or the like becomes unnecessary.

  The advantageous implementation of the method according to the invention, unlike the known list-based method or the bit radius method, processes state information separately for calculation of confidence values, selection of hypotheses and alternative assumptions and their distance metrics. In this case, for example, a correction function necessary for a search based on the MMSE method is incorporated. This method can also be adopted in systems where the distance for calculating the reliability information is different from the distance important for tree search in order to calculate different assumptions and alternative hypotheses by incorporating the correction function. it can. By the optimal calculation of the reliability information using the correction function, the tree search flow with n columns of tapels is not affected.

  That is, for example, in the case of the MMSE method, it is possible to calculate assumptions / alternative assumptions that differ by bias, and furthermore, the search space can be advantageously limited by the MMSE method. As a result, the accuracy of the determined hypothesis and the alternative hypothesis increases, which improves the accuracy of the search. Therefore, this method can be combined with the bit radius method (two lists, one for the bit radius and one for reliability information) and the other tree search methods (width-first search, depth-first search). Furthermore, similarly, the search space calculation can be further optimized without adversely affecting the accuracy of the reliability information. The calculation of the reliability information after the tree search can be performed without further correction processing by subtracting the individual elements from the already adapted state information for each bit. More accurate assumptions and alternative assumptions can be calculated if the stored bit-by-bit metric already includes a correction term for calculation of reliability information (eg, as required by the MMSE method) Thus, the present method can be adopted for a more efficient search algorithm. Thereby, the method is clearly superior to all known methods in terms of efficiency (search accuracy for the analyzed tree nodes). Furthermore, unlike the conventional list-based method, the reliability information can be calculated in parallel by storing each bit.

  Independently of the tree search, by separately processing the state information for calculating the reliability information, the two components of the method can be optimized independently of each other. That is, the calculation of the state information used for limiting the search space can be performed by a method different from the calculation of the state information used for calculating the reliability value. For example, the MMSE method can be used for tree search, and the ZF (zero forcing) method can be used for calculation of state information.

  In another advantageous embodiment of the method, information about the structure of the tree nodes, for example mapping information, can be incorporated when processing the state information for calculating the confidence value separately. To select tree nodes that are considered promising for state information for computing reliability information or tuples for limiting the search space, for example by incorporating information about the structure of tree nodes such as bit mapping The operations for calculating, selecting and storing promising data need only be performed on a possible subset of tree nodes to consider.

  In the present invention, the state information processing for calculating the reliability information can be performed in parallel with the search space determination and the tree search. As a result, the tree search throughput is improved.

  The use of the method according to the invention greatly simplifies the complexity of the tree search method application and at the same time increases the flexibility. As a result, the power efficiency and area efficiency when applied to hardware are significantly increased compared to conventional application forms. By reducing the size of the list, in addition to the memory capacity required for that purpose, the number of operations and the scale required for management are reduced, and at the same time, the complexity of the radius for each bit is reduced. Furthermore, the flexible determination of the search radius allows the search accuracy and the corresponding search load to be adapted to the requirements of the surrounding environment.

  Further advantages and features of the method according to the invention will become apparent from the detailed description of the embodiments given in conjunction with the accompanying drawings.

System model diagram of a MIMO system that iterates the detection / decoding process 64QAM symbol layout of bit value by gray labeling Possible transmission symbol diagram distorted by system matrix Example of tree search for 4x4 MIMO system Data flow diagram of conventional sphere detection algorithm Tuple principle diagram for determining the search space by a subset of the sorted list Principle diagram for storing status information for each bit in a 24-bit example, and storing a distance metric separately for each bit value and each bit Diagram of a variation that saves state information bit by bit in the 24-bit example and distance metric for each bit of hypothesis and alternative hypothesis Variation diagram of standard implementation of list sphere detection FIG. 5 is a diagram illustrating an embodiment of a list sphere detection algorithm for optimally processing state data separately for calculating reliability information according to the present invention. Influence of bias on position determination in the case of MMSE detection method and Q-QAM and correction chart of the influence Comparison of performance of tree search method using depth-first search of the present invention and conventional application Figure of the effect on the performance of the algorithm that stores the optimal state information for calculating the L value for each bit Comparison diagram of application form of tree search method by depth-first search of the present invention and complexity of conventional application form Impact diagram on the complexity of the algorithm that stores the optimal state information for calculating the L value bit by bit Variations of the receiving configuration of FIG. 1 for list-based iteration.

  Here, separation of search space determination from calculation of reliability information according to the present invention and its effect on the tree search method will be described in detail based on examples.

  By separating the search space determination according to the present invention from the reliability information calculation, both the search space determination and the reliability information calculation are optimally performed and the overall complexity of the tree search is greatly reduced. It is possible.

  As already explained in detail at the beginning, in order to calculate the state information according to Equation 1, a series of promising alternative assumptions must be detected in addition to the assumptions during the tree search for detecting the transmission signal.

In the present invention, tuples of n promising tree leaves c ⁱ , not data calculated to calculate reliability information

By limiting the search space for tree search only, sufficiently accurate tree search is guaranteed. In this case, the tuple size n is critical in terms of the quality of the state information (eg distance metric) and hence the complexity of the tree search. By limiting the search space with such tuples in a manner that is separate from the calculation of reliability information, the elements of those tuples are essentially independent of the state information used to calculate reliability information. Become. This is especially true for node selection and status information formats. In this case, the determination of the search space is appropriate for the purposes of the present invention, as follows:

Where element c ⁱ is the detected promising tree leaf,
Is done by. Therefore, according to the progress of the search, the n leaves with the smaller distance metric are always saved temporarily, and the new leaves are compared with the already saved leaves. As a result, the tree search detects n leaves having a smaller distance metric.

  In particular, when the state information calculation method for determining one search space and the other reliability information calculation method are different from each other, the quality and scale of the tuple data are not necessarily the state information for calculating the reliability information. No longer directly related to quality. However, the complexity determined by the tuples and the concentration of the tree search on promising nodes increases the likelihood of finding promising state information for computing reliability information depending on the size of the tuple and the complexity of the search. As a result, the data quality has a high correlation, and therefore represents a sufficient quality regarding the result of the tree search.

  The calculation of the reliability information is different from the known method in any case in order to calculate the reliability information calculated from the distance metric, not the distance metric particularly used for limiting the search space at the time of tree search. This is done based on metrics that can be optimized. Therefore, in the present invention, the calculation of the reliability information (L value) is changed as follows.

Here, f (•) is an optimization function to be used, and σ ² and c are other representative examples of the function necessary for the bias reduction type MMSE detection method which will be described separately with reference to FIG. Is an element. In this case, the optimization function does not apply to the signal candidate list or to all leaves found within the tree search, not just the subset of analyzed leaves that is a set of distance metrics stored by bit radius. Since it is applied, the accuracy of the reliability information is improved accordingly. In particular, the signal candidate c ⁱ found by it that maximizes the metric for c _l = + 1 or c _l = −1 does not agree with the assumption or alternative hypothesis detected by the minimum distance metric.

An advantageous implementation applying the search space determination with n columns of tapels separated from the calculation of reliability information is illustrated in FIG. In this case, a subset (size L> n) obtained by sorting a set of storage elements (6-1, 6-2,..., 6-L) referred to as a list is used as a n-column tapel in the following. . R _Tupel represents the last element of the n-column tapel, the elements of the list up to R _Tupel are sorted according to their distance metric, which is represented by an arrow in the drawing, and R _Tupel is The tuple element with the largest distance metric is shown. The search space for tree search is determined based only on list elements that are constituent parts of n columns of tapels. The remaining elements of the list serve to preserve the alternative assumptions found, and the list is used as a whole for calculating the L value. The new element (found tree leaf) is incorporated into list position 6- (n + 1) with respect to its distance metric (λ ₁ ). Therefore, by incorporating such elements into a list sorted up to the nth element, the first n elements (6-1,. It is equivalent to. Changes in the position of the nth list element after a new leaf has been incorporated, ie, lists 6-1,. . . , 6-L, the size of the tuple (1 ≦ n ≦ L) can be changed by changing the position of the list element 6-n and sorting the list up to that element. Thereby, it is possible to variably determine the size of the tuple and hence the complexity or accuracy of the search, so that the tree search can be adapted to the ambient conditions. As n increases, the number of tree nodes to be searched and hence the complexity of the tree search method increases, but the accuracy of the calculated reliability information also improves. Therefore, the accuracy can be flexibly adapted for each tree search. The search space can be determined by the tuple maximum distance metric by executing a search in which the distance metric of the nth list element is equal to the search radius (

). As described above, starting from the nth element and incorporating a new tree leaf, for example, by sequentially performing a comparison of n elements, the first time (between the nth and n + 1th elements) The search space can be determined immediately after the comparison. Any further comparison certainly serves to sort the list, but does not affect the factors that determine the search space. As a result, in addition to the quick determination of the search space and therefore a short processing time, it is also possible to parallelize operations (n-1 comparisons) for sorting the lists, as will be described in detail separately in connection with FIG. As a result, such an operation does not affect the processing time of the search algorithm.

In another advantageous implementation in which the search space determination by n-column tapels is applied separately from the calculation of the reliability information, it can be adopted independently of the search method used in all tree search methods. By storing the metric bit by bit, the metric f (Λ (c, y, L _a (c)), σ ² , c) optimized for L value calculation based on Equation 4 can be stored separately. Done. In this case, the state information regarding each bit value (0/1) can be temporarily stored separately. FIG. 7 illustrates possible bit-by-bit distance metric storage for 64-QAM and four transmit antennas. In an alternative implementation, state information regarding the bit values of hypothesis c ^h and the alternative hypothesis c _j ≠ c ^h _j resulting therefrom is also present, ie, all bits (0 ... 24) as illustrated in FIG. Hypothetical distance metrics and corresponding bit values and alternative hypothetical distance metrics can be temporarily stored. The newly calculated distance metric is compared to the stored one and always remains more promising, so that after the tree search, the promising metric found as the search progresses is Each can be used for calculation of reliability information. This prevents promising metric losses in response to search progress, such as may occur due to signal candidate list limitations or adverse optimizations in a well-known manner. In addition to the best achievable quality of metrics, the small complexity of the method is also guaranteed by immediately incorporating optimization in combination with bit-by-bit storage. In this case, the required number of memory capacity is small, and each system is set in advance independently of the tree search method, and the size of the list for limiting the search space is given by a given hardware and / or algorithm. It is possible to adapt to the requirements of

  Further, all operations for metric processing and storage can be performed in parallel with the tree search, so that the processing time of the tree search is not hindered by those steps. The computation of reliability information is done directly with individual bit assumptions and alternative assumption metrics, suggesting that the post-tree search complexity is minimized.

A variation of the data flow diagram of the list sphere detection algorithm shown in FIG. 5 is shown in FIG. It has a standard data flow and is described in detail in the applicant's previously filed US Pat. After initialization (step 902), a standard tree search algorithm is performed as follows.
1. If the search has not yet been completed (decision 940), i.e., if the search tree has not been completely searched, or if a stop condition such as the maximum number of clocks has not been met, then continue this algorithm or If there is not, the following 6. The process is executed.
2. As the next step, we will further consider the in-plane child nodes. In this case, as described in detail in Patent Document 2 previously filed by the same applicant, the parent node to be considered next is always known based on initialization or another aspect and determination of the next node. The resulting interference and other previously estimated symbol interference are removed from the received symbol (step 904) and the distance to the child node is calculated (step 905). In this case, for selection of the most probable node, for example, calculation of all child nodes and sorting based on the likelihood (step 906) are necessary.
3. The next step is to investigate whether the search space can be limited (step 914), which is a candidate signal associated with the search sphere, independent of the tree plane being considered at the time. Are stored and sorted (step 912). The sorting criterion is, for example, the reliability of signal candidates.
4). The next plane to consider can then be calculated (step 924).
5. Above 1. Continue to search for trees.
6). The search is terminated (step 940), that is, if, for example, the search tree has been completely searched or the stop condition such as the maximum number of clocks is satisfied, the reliability information is calculated from the detected leaf node. And save (step 942).

  In existing modules, several disadvantages of the well-known search space limiting method or L-value calculation method are exposed. That is, the search space is determined and the L value is calculated in order in the algorithm. In particular, when using a bit-by-bit radius, a series of comparisons must be performed in turn, which is a major drawback in such cases. There, it must first be examined whether the determined leaf represents a new assumption, and if not, whether one or more new alternative assumptions are given, and finally the burden is In this way, the search space must be determined by the bit radius. When using the signal candidate list, necessary sorting is also a problem in addition to the hardware complexity already considered. It can certainly be executed in parallel with the algorithm to some extent, but it becomes a limiting factor that depends on the number of required operations. The L value calculation (step 942) can be executed only after tree search. In some of these methods, in addition to the very high complexity of the algorithm and the corresponding increase in processing time, the L value can only be calculated with stored data, as already explained in detail. It is.

  An advantageous variation of the tree search algorithm with list sphere detection incorporating separately the search space limitation and signal candidate processing for L value calculation according to the present invention is illustrated in FIG. In this case, modules having the same or similar functions as those in FIG. 9 are represented by the same reference numerals increased by 100, and the corresponding data flow is not repeated. After the selection of the next element to be considered (steps 1024 and 1032), only one node on the surface of the tree and the next node to be considered are considered by calculation of the distance metric. Subsequent sorting (step 912 in FIG. 9) has been omitted, and in some cases the inclusion in the radius list has been reduced as an inclusion in the tuple that determines the search space for one leaf node, which is the leaf assumption. Yes. To limit the search space, n columns of tapels are used as described in connection with FIG. By doing so, the search space determination (step 1014) can be made as soon as possible after the first comparison between the nth list element and the new node. Furthermore, another incorporation process can be performed in parallel with the tree search (step 1050) as it does not further change the search space. As a result, the processing time for search space determination (step 1014) is reduced to a minimum. Furthermore, if the search space is determined based solely on leaf assumptions, so that quality and search complexity are the same, only a half-sized tuple is needed. In this case, the calculation of the metric of the occasional leaf and its alternative hypothesis, optimized for the calculation of reliability information, is also in parallel with the search algorithm since such calculation does not affect the search progress. Can be done. Furthermore, with a corresponding selection based on the mapping characteristics, the calculation of alternative assumptions can only be made inexpensively since only for the promising alternative assumptions. With respect to the signal candidate list for calculating the L value, it is proposed to store the signal candidate or state information according to Equation 4 in a form suitable for the detection method. In this case, for both hypotheses and alternative hypotheses, the adverse effects of the search method on the selection of promising (alternative) hypotheses and their metrics, eg, the bias of the MMSE detection method, as will be discussed separately in connection with FIG. The effect can be eliminated. Subsequent to such calculation, the aforementioned bit-by-bit storage is performed according to the metric obtained as a result. Therefore, a metric for calculating the confidence value is obtained as follows:

Therefore, the calculation can be performed by inexpensive subtraction, and the processing time after tree search is minimized.

  By separating the search space determination from the reliability information calculation according to the present invention, different search criteria can be used during tree search than when calculating hypotheses or alternative assumptions and their associated L values. For example, when using the MMSE method for detection, i.e. for tree search, the search results and the resulting reliability information are misaligned, so-called biases, and are not as accurate as expected. . It is caused by noise suppression that minimizes the mean square error. In the case of tree search, the MMSE detection method can be performed by extending the channel matrix prior to QR analysis, for example as described in Non-Patent Document 13, thereby biasing the Euclidean distance.

But as follows:

Will be included. The bias advantageously limits the search space during searching, but introduces errors that must be corrected for accurate detection when calculating reliability information. It generates different metrics for the calculation of reliability information and can also result in different assumptions and alternative assumptions. FIG. 11 (1) illustrates the effect of bias on the metric and the resulting drop in distance. The farther the symbol placement point is from the origin, the greater the reduction in distance due to bias. For this reason, a term with reduced bias is used for the calculation of reliability information.

Must be saved. Thus, in addition to removing the bias from the distance metric for such calculations, changes to assumptions and hence alternative assumptions are also required for reliability information. As shown in FIG. 11 (2), by incorporating such symbol placement points and the influence on the determination region, the geometric consideration in the above-described MMSE detection method is as follows.

Can be migrated to a system compressed with Such disadvantageous compression into the system incurs additional costs during computation. Thereby, as illustrated in FIG. 11 (3), it is more advantageous to shift the influence of the bias to the position representing the search argument. To achieve this, a correction term for calculation or node determination is used. Therefore, to determine the tree face assumption or alternative assumption, use the modified position of the search argument as follows:

The original grid is used for position determination. Regarding the MMSE detection method described above, the correction term obtained from reflecting the signal with reduced bias in the L value calculation is as follows.

In this case, σ ² is exactly the noise variance, and the correction term (1−σ ² ) ⁻¹ corresponds to a compression that is not as accurate as expected due to the reduction in bias. Thereby, the selection of promising leaf nodes can be adjusted. In other words, the expected value of the MMSE detection method is different from the expected value of the bias reduction detection method that matches the expected value of the detection method based on zero forcing. Although this correction term eliminates it, it is only effective for the selection of promising elements. This correction term can be determined before the tree search. Since the assumptions or alternative assumptions for calculating the reliability information and the determination of its distance metric are independent of the original search and search space limitations, this method basically affects the search progress. Absent. However, such improved calculation of reliability information improves the accuracy of the search.

  A good measure for determining the computational complexity and therefore the achievable throughput is the number of node expansions of the tree search.

FIG. 12 shows the performance (BER: bit error rate), and FIG. 14 shows a reduction in search space determination and bias by the above-described 8-row tuple (n = 8), compared with the conventional application form. The complexity of the extended sphere / detection algorithm (average node expansion count) by storing the metric for each bit is shown. In this case, the conventional method obviously requires a lot of node expansion and is therefore not comparable, so this method by bit-by-bit radius and separately optimized bit-by-bit metric processing (3. By extending only radius clipping (up to 5 times), a lossless MMSE detection method with reduced bias is realized. At the time of detection by the detector / decoder iterative method of T = 4, the number of nodes to be considered is simply doubled by the detection method in which the determination of the search space by the n-column tapel according to the present invention is separated. The performance compared to LSD (R = λ _Hypothese list sphere detector) can be greatly improved.

FIG. 16 illustrates a modification of the reception configuration of FIG. 1 for list-based iteration. The additional part of the iterative detection / decoding unit with reduced complexity is the part that reuses the results calculated during the first detection (at detector 1622) to skip the remaining tree search and proceed Have The signal candidate list 1650 calculated during the first detection can be used as the main component for the next iteration (L value calculation at 1642). The basic idea is that without prior knowledge, the state information is small, ie, a large set of leaf nodes with a small distance metric λ ₁ of the found signal candidates or a small optimized metric f (c), for example. Is to find The prior information that can be used for such iterations mainly changes the set of promising nodes as it increases the distance metric of the non-promising path of the tree search. If the signal candidate list from the first detection is sufficiently large, at least one leaf is included, such as a new MAP (maximum a posteriori) leaf or a promising alternative hypothesis in the neighborhood list. Therefore, a new tree search becomes unnecessary.

  Signal candidate list for all bits obtained by the conventional LSD method shown in FIG. 9, calculation for sorting and saving the elements of the list, calculation of necessary hardware unit, reliability information, search for minimum value of list, etc. This complexity of the LSD method search (number of processing steps, hardware units used, power consumption, etc.) as a whole is significantly greater than the application according to the present invention.

  Compared to the radius-by-bit method, the application according to the present invention can greatly reduce the complexity at the same level of performance. Since the complexity of the application form for each node to be considered is almost the same, even in this case, the application form according to the present invention is clearly cheaper and the processing time is significantly shortened.

  Further, FIGS. 13 and 15 illustrate 16 for the case of no such bit-by-bit processing and no detector / decoder iterations, to illustrate the optimal isolated bit-by-bit performance. The performance and complexity of tree searching with row tapels is illustrated. The number of tree nodes to be considered is almost the same as the case of 8 columns of tapels and bitwise storage (uses leaf assumption to limit the search space and halves the size of the tapels). Therefore, it is better than the bit-by-bit radius method omitted in the graphs of FIGS. 13 and 15 for the sake of clarity. However, by storing bit by bit, as already discussed, the complexity of the application according to the present invention is significantly smaller than the method based on the signal candidate list searched by 16 columns of tapels. In addition, there are obvious differences in performance. The performance is improved by about 0.3 to 0.5 dB at the same complexity, compared to comparable search algorithms, just by preventing the optimized processing and hence the loss of promising state information. Therefore, it is a precondition for efficient tree search that the search space is determined and the state information for calculating reliability information is separately and optimally processed. The application of detecting by searching with a suitable size (for example, 16/64) of the n-column taple and the signal candidate list of the present invention enables a nearly optimal detection quality with greatly reduced complexity. It becomes. In addition to improving performance with a corresponding correction function, the complexity of the search can be greatly reduced. This is realized in particular by improving the accuracy of information for reliability calculation and being able to limit the search space independently of the information. By doing so, it is possible to greatly reduce the complexity of the search and to adapt flexibly to ambient conditions.

  A synchronous transfer architecture (STA) can be used to apply the method according to the present invention. By executing the process for each computing unit, the components of the algorithm can be integrated accordingly into one or more STA modules. The modules can be controlled individually or by one VLIW. Thereby, the advantages of STA, high performance, low power consumption, simple expandability and high flexibility can also be utilized for the tree search method.

  The present invention can be used in combination with an OFDM (Orthogonal Frequency Division Multiplexing) system as a digital transmission method. In that case, the system matrix H has a transmission channel in the frequency band for one or more subcarriers. The present invention can also be used in combination with other multiple access schemes such as direct sequence CDMA (DS-CDMA), multicarrier CDMA (MC-CDMA), and space division multiple access (SDMA).

  The present invention can be applied to one or a plurality of receivers that jointly detect received data even in a multi-user transmission (MUT) scheme. In that case, the system matrix H has a transmission channel across the user. Multiple antennas can be used in one or more receivers.

  From the above description, in the tree search method in which the search space determination proposed here is separated from the calculation of the reliability information, the search radius can be easily determined, the number of necessary storage elements is reduced, and the search space is determined. For example, when the performance is the same, the complexity of the algorithm is greatly reduced compared to known methods, such as reducing the burden of sorting the tapels used for the search. It has become clear that efficient hardware applications with reduced complexity are possible, which makes it possible to adopt the method in an actual mobile system.

Claims (14)

A tree search method based on depth-first search for detecting a received signal of a multi-input / multi-output (MIMO) channel based on a tree search, wherein a leaf node of the search tree represents a possible transmission bit and has the highest likelihood. And the reliability value of that bit is determined based on the state information of the leaf nodes detected during the tree search, and the search space for the tree search is determined. Is limited based on leaf node state information detected during tree search,
Limiting the search space for this tree search is performed independently of the processing of state information for calculating confidence values;
Limiting this search space is done by n columns of tapels in the leaf node state information detected during the tree search;
The maximum value of the state information in these tapels is determined as the radius of the search space;
A method characterized by.   During a tree search, leaf nodes with more promising state information than the least promising value of the n-column tapel, found within the current search radius, are taken into the n-column tapel and its most The method of claim 1, wherein the method replaces a non-promising element.   2. The n columns of tapels are generated as a sorted list of list elements 1 to n, and the list elements are arranged according to state information values. the method of. The n columns of tapels are implemented as a subset of a list of size L where L> n, and this list has L pieces of state information for calculating confidence values;
The list of size L is sorted to at least element n + 1;
The method according to claim 3.   The method according to any one of claims 1 to 4, wherein the state information is a distance metric of a leaf node of a search tree.   To limit the search space, the nth list element of the list of size L is used, and the position of n in the list of size L is variable, where 1 <n ≦ L. 6. A method according to any one of the preceding claims, characterized in that it is characterized in that State information used to calculate the confidence value is stored separately from the n-column tapels used to limit the search space;
The state information does not affect the search flow of the tree search;
The method according to claim 1, characterized in that:   The method according to claim 7, wherein storing the state information for calculating the reliability information separately is performed by storing the information for each bit.   9. The method of claim 8, wherein storing the state information for each bit is performed separately for each possible bit value and for all bits.   Saving the state information bit by bit is performed in common for all the bits of the current hypothesis for a single transmitted symbol, and separately for the corresponding alternative hypothetical bits. 9. The method of claim 8, wherein the method is performed.   11. A method according to claim 1, wherein a correction function is incorporated in the separate processing of the state information for calculating the reliability information. For the tree search, the MMSE (Minimum Mean Square Error) detection method is used, and for the calculation of the tree plane assumption and the alternative assumption, the corrected position of the received signal is used as a correction function of the MMSE detection method. The method of claim 11, wherein a correction term (1-σ ² ) ⁻¹ is used, where σ ² is noise variance.   The information regarding the structure of the tree node, for example, mapping information is incorporated when separately processing the state information for calculating the reliability value, according to any one of claims 1 to 12. The method described.   The method according to any one of claims 1 to 13, wherein the processing of the state information for calculating the reliability information is performed in parallel with the search space determination and the tree search.

Images