3. Dynamic Cell-Less Architecture

To tackle the issues of fronthauling overhead and high complexity of massive cellless networks, some recent works have proposed signal processing schemes that distribute the pre-processing and detection functionalities among CPU pools and APs while maintaining a static property of the network by assigning identical tasks to all APs (i.e. all APs are selected to take part in the processing of signal detection in an identical manner). This can be referred to as multi-level AP cooperation and can be achieved through different signal processing schemes (maximal ratio combining [MRC], minimum mean square error [MMSE], zero forcing beam-forming [ZFBF], etc.) [9]. Such a distribution of tasks between CPU Pools and APs will result in a trade-off between performance and system complexity. Building on top of the above concept, we propose a dynamically changing cellless architecture that can significantly decrease the amount of fronthualing overhead and central processing by updating the allocation of tasks among the APs according to the instantaneous channel state information (CSI) of the network. The idea of the proposed scheme is that the set of serving APs is dynamically partitioned into subsets of APs clustered based on the current CSI of the network (see Fig. 2). Furthermore, all APs of a certain subset act as a virtual collocated transmitter(receiver) diversity (V-TRD) system [10, 11]. Therefore, the CPU will deal with every subset as a single virtual AP with multiple antennas. Due to the decrease in the number of real APs directly connected to the CPU pool, this will reduce the fronthauling overhead and optimization variables significantly. Specifically, compared with (1), the SINR for the k-th user at the CPU detector unit can

comes complicated and time consuming. Additionally, when the number of sub-sets becomes small, performance degradation compared to the fully centralized networks can increase significantly. Note that a small number of subsets means that most of the APs will be used as antennas for the VTR-Diversity system. The maximum number of subsets may be equal to the total number of APs in the net-work such that every subset contains only one AP. This scenario is identical to the fully-centralized network scenario. The CPU deals with every subset as a sin-gle virtual AP with multiple antennas. Therefore, processing is more distributed, especially when the number of APs per set is large (i.e. the number of AP sets is small). In such a scenario, the CPU will have a fewer degrees of freedom in optimizing the performance (i.e. shorter encoding/precoding vector length). As an example, if we have only two subsets, all of the APs will be allocated between those two sets. The CPU then will have to control only a weighted sum of two signals coming from these two subsets. However, if we have four subsets, the CPU will have more degree of freedom since it will optimize a weighted sum of four signals rather than only two in the previous case.

In the following two subsections, we first discuss how clustering can be done based on some intelligence(AI)-based technique. Furthermore, we also propose the utilization of NOMA scheme to improve spectral efficiency of the network.

3.1.AP Clustering

The process of partitioning APs into different subsets can be modeled as an opti-mization problem that allocates the APs among a predetermined number of sub-sets(the number of subsets is selected based on factors such as the required reduction level of complexity, fronthauling overhead, and/or delay). However, solv-ing such an optimization problem may be very complicated and time consuming, especially for massive numbers of APs and/or terminal users. This can be tack-led by investigating a group of sub-optimal clustering algorithms that can be easily solved with minimum possible delay. One example of these algorithms is to assign equal number of APs to every subset and allocating every subset with a group of APs that have the maximum norm with any arbitrary terminal user. This strategy may be considered as a simplified MRC scheme that focuses on enhancing the user’s signal related to the best channel gain with a certain subset. Usually, for distributed MIMO systems, the receiver first uses MRC technique as the first step of decoding an intended signal. This is achieved by multiplying the received signal from every AP by the complex conjugate of the channel gain between the signal coming from the intended user and that AP. In the proposed scheme, we multiply the overall signal of every subset (the sum of received signals from all APs at a certain subset) by the complex conjugate of the sum of channel gains of the intended user and all APs of that subset. The performance of such an allocation algorithm can be very poor compared to that of the optimal one.

To overcome the complexity of exhaustive search-based optimal clustering, ar-tificial intelligence (AI)-based methods can be developed to obtain an efficient allocation of APs among different subsets (in terms of maximizing the average per user SINR). Fig. 3 shows the block diagram of such an AI-based approach (which uses deep reinforcement learning in particular) that accepts the channel gain matrix (denoted by H), number of subsets of APs, and the number of APs per each subset, and returns an assignment of every AP in the network to a certain subset.

3.2.NOMA-Enabled Dynamic Cell-Less Architecture

We have showed in previous sections how moving from fully-centralized cell-less architectures into dynamic cell-less networks with partially distributed processing significantly decreases the system complexity and delay. However, such a design can result in a degradation of network performance as the number of AP subsets and/or the number of terminal users decreases. To compensate for this perfor-mance degradation, we propose the use of NOMA scheme among all terminal users under full cooperation among APs [1]. Accordingly, assuming that all termi-nal users are ordered in a descending ordering, the SINR of the k-th user at the