Supélec, 2012-02-02 Optimization Concepts for Resource Allocation in Cellular Systems Emil Björnson PhD in Telecommunications Signal Processing Lab KTH Royal Institute of Technology KTH in Stockholm KTH was founded in 1827 and is the largest of Sweden’s technical universities. Since 1917, activities have been housed in central Stockholm, in beautiful buildings which today have the status of historical monuments. KTH is located on five campuses. 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 2 A top European grant-earning university • Europe’s most successful university in terms of earning European Research Council Advanced Grant funding for ”investigator-driven frontier research” 5 research projects awarded in 2008: • Open silicon-based research platform for emerging devices • Astrophysical Dynamos • Atomic-Level Physics of Advanced Materials • Agile MIMO Systems for Communications, Biomedicine, and Defense • Approximation of NP-hard optimization problems 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 3 Emil Björnson • 1983: Born in Malmö, Sweden • 2007: Master of Science in Engineering Mathematics, Lund University, Sweden • 2011-11-17: Defended doctoral thesis in telecommunications, KTH, Sweden Multiantenna Cellular Communications Channel Estimation, Feedback, and Resource Allocation – Three Building Blocks of Physical Layer – Mathematical Analysis and Optimization 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 4 Background • Cellular Communications - Many transmitting multi-antenna base stations - Many receiving single-antenna users • Downlink Transmission - Multiple transmit antennas – exploit spatial dimension - Multiuser transmission Pro: Higher performance, Con: Co-user interference 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 5 Background (2): Multiple Cells • Uncoordinated Cells: + + − − Simple processing Simple infrastructure Uncontrolled interference Or fractional frequency reuse • Coordinated Cells: + − − − 2012-02-02 Controlled interference Backhaul signaling Computationally complex Tight synchronization Emil Björnson, KTH Royal Institute of Technology 6 Background (3): Multiple Cells • Dynamic User-Centric Coordination Clusters - Inner Circle (Strong Channels): Consider Transmitting to Users - Outer Circle (Non-negligible Channels): Avoid Interference to Users - Can model any level of coordination 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 7 Background (4) • Resource Allocation - Select users for transmission - Design beamforming directions - Allocate transmit power • Optimize Resource Allocation - Maximize system performance Satisfy system constraints (power, interference, fairness) Any level of coordinated multipoint transmission Robustness to uncertain channel information • No mathematical details - Focus on performance optimization concepts - Assumption: Linear transmit/receive processing 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 8 Outline • What is Performance? - Different user performance measures - System Performance vs. user fairness • Multi-user Performance Region - How to interpret? - How to choose operating point? • Performance Optimization - Geometrical interpretation of common formulations - Right problem formulation = Easy to solve • Low-complexity Strategies - Exploit structure from optimal solution 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 9 What is Performance? 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 10 What is Performance? • Service Quality - Experienced by users (per-user level) - Can also be measured at system-level • Performance Based on - Average data rate Latency Coverage Battery life Etc. • Simplified Performance Measures - Necessary for optimization 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 11 Single-user Performance Measures • Mean Square Error (MSE) - Difference: transmitted and received signal - Easy to analyze - Far from user perspective? All improves with SNR: • Bit/Symbol Error Rate (BER/SER) - Probability of error (for given data rate) - Intuitive interpretation - Complicated & ignores channel coding Signal Power Noise Power Optimize SNR instead! • Data Rate - Bits per ”channel use” - Mutual information: perfect and long coding - Still closest to reality? 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 12 Multi-user Performance • User Performance Measures - Same measures – but one per user • Performance Limitations - Power Allocation - Co-user interference: SINR= Signal Power Interference + Noise Power • Why Not Increase Power? - Power = Money & Environmental Impact - Reduce noise Interference limited • User Fairness - New dimension of difficulty - Heterogeneous user conditions - Depends on performance measure 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 13 Multi-user Performance Region 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 14 Multi-user Performance Region • Achievable Performance Region – 2 users Performance User 2 - Under power budget Care about user 2 Balance between users Part of interest: Upper boundary Performance Region Care about user 1 Performance User 1 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 15 Multi-user Performance Region (3) • Can it have any shape? • No! Can prove that: - Compact set - Simply connected (No holes) - Nice upper boundary Normal set Upper corner in region, everything inside region 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 16 Multi-user Performance Region (3) • Possible Shapes of Region - Convex, concave, or neither - In general: Non-convex - In any case: Region is unknown Convex 2012-02-02 Concave Emil Björnson, KTH Royal Institute of Technology Non-convex Non-concave 17 Multi-user Performance Region (3) • Some Operating Points – Game Theory Names Performance User 2 Utilitarian point (Max sum performance) Egalitarian point (Max fairness) Single user point Performance Region Which point to choose? Optimize: Sum Performance? Fairness? Single user point Performance User 1 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 18 Performance Optimization 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 19 System Performance versus Fairness • Always Sacrifice Either - Sum Performance - User Fairness - Or both: optimize something in between • Two Standard Optimization Strategies Starts from Performance Starts from Fairness - Maximize weighted sum performance: maximize w1·R1 + w2·R2 + … R1,R2,… (w1 + w2+… = 1) - Maximize performance with fairness-profile: maximize Rsum Rsum subject to R1=a1·Rsum, R2=a2·Rsum, … (a1 + a2+… = 1) • Non-Convex Optimization Problems - Generally hard to solve numerically 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 20 The “Easy” Problem • Given Point (R1,R2,…) - Find transmit strategy that attains this point - Minimize power usage • Convex Problem - Second-order cone or semi-definite program - Global solution in polynomial time – use CVX, Yalmip Single-cell (total power) • M. Bengtsson, B. Ottersten, “Optimal Downlink Beamforming Using Semidefinite Optimization,” Proc. Allerton, 1999. • A. Wiesel, Y. Eldar, and S. Shamai, “Linear precoding via conic optimization for fixed MIMO receivers,” IEEE Trans. Signal Processing, 2006. Single-cell (per ant. power) • W. Yu and T. Lan, “Transmitter optimization for the multi-antenna downlink with per-antenna power constraints,” IEEE Trans. Signal Process., 2007. Multi-cell (general power, robustness) • E. Björnson, G. Zheng, M. Bengtsson, B. Ottersten, “Robust Monotonic Optimization Framework for Multicell MISO Systems,” IEEE Transactions on Signal Processing, To appear. 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 21 Exploiting the “Easy” Problem • Easy: Achieve a Given Point • Hard: Find a Good Point Main part of resource allocation • Shape of Performance Region - Far from obvious – one dimension per user Interference Channel Rate: user 1 3 transmitters w. 4 antennas 3 users Rate: user 2 2012-02-02 Rate: user 3 Emil Björnson, KTH Royal Institute of Technology 22 Geometric Optimization Interpretations • Maximize Performance with Fairness Profile: maximize Rsum Rsum subject to R1=a1·Rsum, R2=a2·Rsum, … (a1 + a2+… = 1) • Geometric Interpretation - Search on line in direction (a1,a2,…) from origin (a1,a2,…)·Rsum =(a1·Rsum,a2·Rsum,…) 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 23 Geometric Optimization Interpretations (2) • Simple line-search algorithm: Bisection - Non-convex Iterative convex (Quasi-convex) 1. Find start interval 2. Solve the “easy” problem at midpoint 3. If feasible: Remove lower half Else: Remove upper half 4. Iterate Subproblem: Convex optimization Line-search: Linear convergence One dimension (independ. #users) 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 24 Geometric Optimization Interpretations (3) • Maximize weighted sum performance: maximize w1·R1 + w2·R2 + … R1,R2,… (w1 + w2+… = 1) • Geometric interpretation - Search on line w1·R1 + w2·R2 = max-value Max-value is unknown! - 2012-02-02 Distance from origin unknown Line hyperplane (dim: #user – 1) Harder than fairness-profile problem! Iterative search algorithm? Emil Björnson, KTH Royal Institute of Technology 25 Geometric Optimization Interpretations (4) • Systematic Search Algorithm - Concentrate on important parts of performance region - Improve lower/upper bounds on optimum: - Continue until • Efficiently Solvable Subproblems - Based on Fairness-profile problem 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 26 Geometric Optimization Interpretations (5) • Branch-Reduce-Bound (BRB) Algorithm 1. 2. 3. 4. Cover region with a box Divide the box into two sub-boxes Remove parts with no solutions in Search for solutions to improve bounds (Based on Fairness-profile problem) 5. Continue with sub-box with largest value 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 27 Geometric Optimization Interpretations (6) Properties - Global Convergence - Accuracy ε>0 in finitely many iterations - Exponential complexity only in #users - Polynomial complexity in other parameters (#antennas/constraints) 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 28 Geometric Optimization: Conclusions • Fairness-Profile Approach: Easy - Quasi-Convex: Polynomial complexity - Reason: Only one search dimension • Weighted Sum Performance: Difficult - NP-hard: Exponential complexity (in #users) - Reason: Optimizes both performance and fairness • Every Weighted Sum = Some Fairness-Profile - Easier to solve when posed as fairness-profile problem - Parameter relationship non-obvious 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 29 Geometric Optimization: References • Line-Search Algorithm for Fairness-Profiles • M. Mohseni, R. Zhang, and J. Cioffi, “Optimized transmission for fading multiple-access and broadcast channels with multiple antennas,” IEEE J. Sel. Areas Commun., vol. 24, no. 8, pp. 1627–1639, 2006. • J. Lee and N. Jindal, “Symmetric capacity of MIMO downlink channels,” in Proc. IEEE ISIT’06, 2006, pp. 1031–1035. • E. Björnson, M. Bengtsson, and B. Ottersten, “Pareto Characterization of the Multicell MIMO Performance Region With Simple Receivers,” IEEE Trans. on Signal Processing, Submitted, 2011. • BRB Algorithm - Useful for more than weighted sum performance - E.g. arithmetic, geometric, or harmonic mean performance • H. Tuy, F. Al-Khayyal, and P. Thach, “Monotonic optimization: Branch and cut methods,” Essays and Surveys in Global Optimization, Springer, 2005. • E. Björnson, G. Zheng, M. Bengtsson, B. Ottersten, “Robust Monotonic Optimization Framework for Multicell MISO Systems,” IEEE Transactions on Signal Processing, To appear. 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 30 Low-complexity Strategies 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 31 Low-complexity Strategies • Hardware Limitations - Polynomial complexity: Only slowly-varying channels - Exponential complexity: Only suitable for benchmarking • Heuristic Resource Allocation - Find reasonable strategy with little effort - Exploit available insight the optimal structure • Parametrization of Upper Boundary 1. Select parameters in [0,1] 2. Get an strategy explicitly - Can achieve any point on upper boundary - Only necessary condition 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 32 Low-complexity Strategies (2) • Method 1: Interference-temperature Control - Transmitters x (Receivers – 1) parameters • X. Shang, B. Chen, and H. V. Poor, “Multi-user MISO interference channels with single-user detection: Optimality of beamforming and the achievable rate region,” IEEE Trans. Inf. Theory, 2011. • R. Mochaourab, E. Jorswieck, “Optimal Beamforming in Interference Networks with Perfect Local Channel Information,” IEEE Trans. Signal Processing, 2011. • Method 2: Exploit Solution Structure of “Easy” Problem - Explicit strategy given by optimal Lagrange multipliers Always same structure, but different parameters Take Lagrange multipliers as our parameters! Transmitters + Receivers – 1 parameters • E. Björnson, M. Bengtsson, and B. Ottersten, “Pareto Characterization of the Multicell MIMO Performance Region With Simple Receivers,” IEEE Trans. Signal Processing, Submitted, 2011. 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 33 Low-complexity Strategies (3) • Number of Parameters - Large difference for large problems 100 80 60 Method 1 40 Method 2 20 0 2 3 4 5 6 7 8 9 10 Number of Transmitters/Receivers 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 34 Low-complexity Strategies (4) • Parameterized Structure of Resource Allocation - Beamforming directions maximize: Signal Power at User 𝑘 µ𝑗 · Noise Power + 𝑙 λ𝑙 · Interference Power to User 𝑙 - Parameter λ𝑙 = Priority of User 𝑙 - Parameter µ𝑗 = Based on transmit power of Base station 𝑗 • Heuristic Selection: - All Parameters = 1 - Called: Signal-to-leakage+noise ratio (SLNR) beamforming - Known to work well! 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 35 Example 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 36 Example – Multicell Scenario • Maximize Weighted Sum Rate - Two base stations: 20 dBm output power - Full, Partial, or No coordination BRB algorithm 2012-02-02 Heuristic Parameters (=1) Emil Björnson, KTH Royal Institute of Technology 37 Summary • Easy to Measure Single-user Performance • Multi-user Performance Measures - Sum performance vs. user fairness • Performance Region - All combinations of user performance - Upper boundary: All efficient outcomes - Explicit Parametrization: Low-complexity strategies • Two Standard Optimization Strategies - Maximize weighted sum performance • Difficult to solve (optimally – heuristic approx. exists) - Maximize performance with fairness profile • Easy to solve (with line-search algorithm) 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 38 Thank You for Listening! Questions? Papers and Presentations Available: http://www.ee.kth.se/~emilbjo 2012-02-02 Emil Björnson, KTH Royal Institute of Technology 39