Machine Learning for Optimal Resource Allocation.pptx
- 2. To achieve the postulated performance gains for 5G and subsequent generations 1) Exhibit strict quality-of-service (QoS) and network connectivity requirements even for mobile users at the cell edges and under severe interference.. 2) An expansion of the utilized frequencies to the millimeter-wave range and the deployment of massive multiple-input and multiple-output (MIMO) antenna arrays. 3) a third important resource domain has gathered significant attention: an increase in the spatial density of the network architecture. This has been identified as a challenge in the operation of ultra dense wireless networks especially if the corresponding network architecture requires decentralized network control based on local channel state information (CSI) and limited inter-cell coordination Instructions
- 3. Network management and resources allocation in multi-cell depend on binary and integer decision-making Binary decision-making: choose value yes or no exp. When user request a channel allocation Integer decision-making: select value from list of values exp. When multi users comparative for limited resources allocation the selected value represent the amount for resources allocate to the use Binary decision-making Integer decision-making Instructions Fast Simple less accurate Slower Complicated More accurate Need more resources
- 4. As today’s cellular networks are fundamentally limited by interference, the associated integer programming problems are generally of a combinatorial nature where optimization is carried out with the goal of trading off conflicting interests among players in the network ,and solutions are at best locally Pareto-optimal. Pareto-optima is optimal with respect to the trade-offs between conflicting objectives, but it may not be optimal in an absolute sense. This type of solution is often used in complex optimization problems where finding the globally optimal solution is intractable or impractical. Use integer programming problems on small-medium network to learn ML as data entry Instructions
- 5. The key question is how well the machines generalize their knowledge to networks of different sizes, topologies, or underlying channel characteristics?
- 6. Network Capacity and Densification The deployment of (indoor and outdoor) small cells is understood as a necessary supplement of this technology to provide widespread network coverage and capacity increase. there are practical reasons for the limitation of wireless network densification, such as the associated increases in hardware costs and energy consumption as well as limited availability of deployment sites and backhaul . However, the primary reason for this saturation effect in densification is the increase of interference in the network, causing deterioration of the signal-to- noise-ratios (SINR) So network optimization and access control is required to control and balance the resources consumption . Employ load-Balancing between the cells Load : The load is defined for each cell as the ratio of its consumed to its available resources Overloaded cell :has many requests from DPs with limitation in spectral resources and will make the link weak (weakest link) ,solution to reduce the load it to
- 7. Decentralized Resource Minimization ● Finding the allocation that optimizes the entire network operation naturally requires information about all network entities and all possible wireless links. ● Network optimization may be performed online during operation, while allocation rules may be devised beforehand based on network information and demand forecasts. . For predefined bias values, this approach operates fully decentralized, as mobile devices can autonomously select their access point based on received signal strength measurements
- 8. Decentralized Resource Minimization Risk assessment range expansion Social networks Overloaded MC convert the users for Mc to SM with lower power MCs= high power large converge area SCs= low power small converge area range expansion even when SCs have lower power the users goes to it Bias value = the difference between Mc power and Sc power It must gather the bias value from all cell and that causes coordinate overload ML is employed as a tool to provide decentralized solution for the combinatorial network optimization problem. Multi-class support vector machines (SVMs) and artificial neural networks (ANNs)
- 9. System Model In this section, a system model for a heterogeneous wireless network with macro cells (MCs) and small cells (SCs) is introduced. The cells in the network serve multiple demand points (DPs) that may represent single mobile users, aggregated data demand from hotspots, or machine-like entities, e.g. in sensor networks. In practical networks, the allocation of DPs to cells is subject to common predefined allocation rules, which are introduced in the following.
- 10. Heterogeneous Wireless Networks K cells in total C = {1,…, K}. C^MC⊂ C and C^SC ⊂ C with C = CSC C^MC ∪ and C^SC C^MC = ∩ ∅ M DPs, with the set of all DPs M= {1,…,M}. DP m M ∈ dm in bits per second. The transmit power of cell k is denoted as Pk AWGN 𝜎2. gkm is determined by the antenna gains of the base station and the DP antennas, respectively, the path attenuation factor, and the processing gain achieved at the receiver by coherent multiantenna processing schemes such as maximum ratio combining (MRC) and zero forcing (ZF)
- 11. Heterogeneous Wireless Networks . the transmission rate for the wireless link between cell k and DP m is upper-bounded by: Where W total system BW. n^bw is the bandwidth efficiency is the bandwidth efficiency corresponding to the selected modulation and coding scheme. To satisfy the data demands ofDPm, cell k needs to utilize at least the fraction dm∕Rkm of its available resources. The allocation of DPs to cells is in the following indicated by the binary matrix A {0 ∈ , 1}K×M with entries defined as
- 12. Heterogeneous Wireless Networks The resource consumption Φk of a cell k is a measure for the fraction of total resources consumed by the cell to serve the demands of all its allocated DPs. It is defined as A cell k is considered overloaded if the resource consumption exceeds one, i.e. Φk ≥ 1. where the following worst-case interference assumptions are made: (A1) Always active: Cells are assumed to be always active, hence the on/of switching of cells is not considered. transmit control channels even in the absence of user (A2) Full load: we assume that DPs have full buffers and always use all their available resources. A1 and A2 are worst case Using supervised learning-based to prevent A1 and A2 to happens
- 13. Load Balancing A straightforward rule for allocating DPs to cells, referred to as maximum receive power allocation, allocating each DP to the cell that provides the strongest signal. Reduces network performance as it leaves the typically low-power SCs underutilized. Solve by Range expansion. In range expansion, the total received power pk gkm from cell k is scaled with a weighting sum resource consumption Condition Σk c ∈ Akm =1 lowest additional load,
- 14. Resource Minimization Approaches This section discusses the allocation of base stations to data points based on an optimal design of the allocation matrix and bias values for range expansion. A mixed-integer linear program is formulated to minimize network resource consumption while satisfying QoS constraints for each data point. However, this approach is computationally demanding and requires centralized CSI, which is practically intractable for large-scale networks. To address this, a decentralized supervised learning-based approach is considered to obtain close-to-optimal solutions with low complexity and local CSI. Data labels are generated based on the optimal solution of the integer programming problems, and classifiers are trained to replicate the optimization behavior using locally available network information and selective features.
- 15. Optimized Allocation to achieve a minimum SINR threshold 𝛾MIN associated e.g. with a given requested modulation and coding order received power must be greater than the sum of the interference from other cells The MILP is designed to optimize the allocation of DPs to cells while achieving resource minimization . The required amount of resources
- 16. Feature Selection and Training For the proposed learning-based resource-minimization approach, each DP extracts system attributes corresponding to three allocation candidate cells as features. These features must be sufficiently selective for the classification and are chosen according to engineering experience. The first feature in our supervised learning-based approach is a cell-type indicator F^TYPE(k) = { 1 if cell k is a small cell, 0 otherwise. The second attribute describes the additional load that user m would cause to cell k if the user was allocated to it: F^LOAD(k,m) = dm /BW log2 (1 + km 𝛾 ) The third attribute is the sum load that would be caused to cell k by all DPs in its coverage area under max.-receive-power allocation: F^COV(k) =Σdm /BW log2 (1 + km 𝛾 ) m|𝜅Pm=k The feature vector
- 17. Range Expansion Optimization Find optimal bias value that optimize the load on the cell Due to the bilinear product terms Akm k 𝜃 of optimization variables Akm and k 𝜃 in constraint (5.8), the problem (5.19) is a mixed integer nonlinear program (MINLP) for which currently no efficient solvers are available. To solve the problem efficiently with general-purpose optimization software, we apply a lifting technique that converts the problem into an equivalent MILP An auxiliary parameter Δkm is introduced such that Δkm = Akm k 𝜃 ∀k,m holds. This can be enforced by the following set of linear inequalities:
- 18. Range Expansion Classifier Training Number of DPs that connect to k cell. Received power grater than pjgjm threshold GTYPE(k) = 1 if cell k is a SC and GTYPE(k) = 0 The load in primary cell The load in secondary cell This is the vector that learning-base will use For resources allocation
- 19. Multi-Class Classification An outline of how these classification problems can be solved using SVMs and ANNs is provided in the following. Using soft-max :mathematical operation user h vector that content R number into probability distribution ANNs SVMs
- 20. Numerical Results dm = 1Mbit ∕ s ∀ m. The resource consumption again increases mostly linearly with M