4.7 Non-maximum likelihood tree-search based detectors
4.7.2 Hybrid tree-search detectors
4.7.2.1 The Adaptive Reduced Breadth-First Search algorithm
Lai et al. in [84] have examined the possibility of the hybrid tree search. A two stage search was proposed. In the first stage a full-blown breadth-first (FBF) search is performed for NBF levels followed by the sequential DFS tree traversal for every subtree.
After the FBF search the symbol vectors sNN
BF representing the nodes on level NBF are
sorted based on the path metric M (sNN
BF) of the symbol vectors. The sorted symbol
vectors are denoted as sN <0>N
BF , s
N <1> NBF , · · · , s
N <|Ω|NBF>
NBF , where the lowest path metric is
achieved by sN <0>N
BF . For every expanded node s
N <j>
NBF a subtree of depth N − NBF is
associated, referred to as subtree j. Every subtree is traversed using the DFS similar as in the SD algorithm. The traversal starts on the subtree associated to the best path metric node. Whenever the traversal of subtree j is completed, subtree j + 1 is selected. The DFS search is terminated if M (sN <j+1>N
BF ) > d
2 or all the subtrees have been searched. At this point the FBF-DFS algorithm finds the ML solution, the only difference compared to the SD algorithm is that there is a sorting on level NBF and the DFS is performed in
the ascending order of the path metrics.
In [84] the FBF-DFS algorithm was extended and further computational complexity reduction was achieved in the Adapting Reduced BF-DFS (ARBF-DFS) algorithm. The extensions of the ARBF-DFS algorithm are: (i) the channel dependent FBF level NBF and (ii) the reliability index based tree pruning.
In the ARBF-DFS algorithm NBF is allowed to vary indepently 1 ≤ NBF ≤ NBF,max, where Nmax denotes the maximum number of levels allowed for the BF stage. If the
channel condition number is favorable NBF is closer to one, thus, NBF becomes a random variable. The worst case memory requirement is proportional to ∼ |Ω|NBF,max.
The second enhancement was the introduction of the reliability index. The aim of the reliability index is to help in the pruning of the tree, thus, the DFS stage is not
4.7. NON-MAXIMUM LIKELIHOOD TREE-SEARCH BASED DETECTORS
performed on every subtree resulting in complexity reduction. In order to give a deeper insight, the partial ML solution is defined as skM L = (sk, sk+1, · · · , sN). The reliability
index is defined as P (sN <j+1>k = skM L|nk, R), namely the probability of sN <j+1>k being
the correct path given the noise realization nkand R. A straightforward way to determine when to stop the FBF stage is to define a threshold T and evaluate if
P (sN <1>k = skM L|nk, R) > T. (4.48)
Furthermore L<j>k is define as follows:
L<j>k = ln P (s N <1> k = skM L|nk, R) P (sN <j>k = skM L|nk, R) ! = M (s N <j> k ) − M (sN <1>k ) σ2 (4.49)
where the final result is achieved by using the path metric formula and the Gaussian probability density function. Suming the probabilities leads to a different form of the reliability index as follows:
|Ω|k X j=1 P (sN <j>k = skM L|nk, R) = 1 P (sN <1>k = skM L|nk, R) = 1 P|Ω|k j=1e −L<j> k (4.50)
The computation of Eq. 4.50 could be very expensive if k is large, thus, it can be simplified by computing only the first two terms of the sum. In this case the approximation of Eq. 4.50 becomes P (sN <1>k = skM L|nk, R) = 1 P|Ω|k j=1e −L<j> k ≈ 1 1 + e−L<2>k (4.51)
By substituting Eq. 4.49 in 4.51 and rewriting Eq. 4.48 the resulting inequality is
M (sN <2>k ) > M (sN <1>k ) + σ2ln
T
1 − T
. (4.52)
This result can be further generalized, namely every partial symbol vector on levels 1 < k < NBF whose path metric exceed the best path metric with threshold σ2ln T
1−T
are pruned.
The introduced pruning reduces the average floating point operations by 2-5 times compared to the SD algorithm. The BER performance is highly influenced by the chosen threshold T . BER simulations with T = 0.9 achieved near-ML performance.
4.7. NON-MAXIMUM LIKELIHOOD TREE-SEARCH BASED DETECTORS
4.7.2.2 The Fixed-Complexity Sphere Detector algorithm
The Fixed-Complexity Sphere Detector (FSD) was introduced by Barbero et al. in [56]. The FSD algorithm tree traversal is built on a hybrid scheme as well. However, several important differences are introduced compared to the ARBF-DFS algorithm. The main features of the FSD: (i) a channel independent tree traversal process consisting of full BFS on several levels followed by a single DFS path based on the SE enumeration and (ii) a novel channel matrix ordering method.
A conjecture was presented about the number of nodes that have to be visited in an uncoded MIMO system in order to achieve near-ML performance. The result is that two stages are defined: (i) in the FBF stage the maximum number of nodes are considered for
NBF levels and (ii) in the single path DFS stage only one child node is further expanded
for NSElevels and the selected child is based on the SE enumeration. The levels sum of the
two stages NBF+NSE = N equals the depth of the tree. The above configuration implies that the number of predefined paths is fixed and is equal |P | = |Ω|NBF·1NSE. In [85] it was
shown that FSD achieves the same diversity as the ML detection if NBF ≥√M − 1. The
advantage of the predefined paths is the resulting rigid structure that enables the parallel implementation of the FSD algorithm and has the same computational complexity for every SNR value.
Another important part of this algorithm is the channel matrix ordering method. It determines the order of the symbols detection based on the stages of the algorithm. The aim is to detect symbols with the highest post-detection noise amplification in the FBF stage and symbols with the smallest post-detection noise amplification in the DFS stage. The motivation behind this strategy is to keep all the paths where the error probability is high and after that it is enough to follow only a single path from every expanded node because the probability of making a bad decision with the SE enumeration is low. The symbol sk to be detected is selected according to
k = arg max j k(Hi†)jk2 , if i is a FBF level arg min j k(Hi†)jk2 , if i is a DF level
where (Hi†)j denotes the j-th row of Hi† and the calculation of matrix Hi† is detailed in Alg. 1 Sec. 4.4.2.
The FSD achieved near-ML BER performance for smaller systems, however, the dif- ference of the BER is getting bigger as the number of antennas are increasing or channel