Optimal Power Control for Wireless Queueing Networks Tolga Girici and Anthony Ephremides Abstract Optimally controlling transmit power for jointly minimizing: System Model Wireless Link Finite buffer capacity (length= K) Bernoulli packet arrivals (rate= λ) Equal length packets (length= 1 bit) Energy expenditure Buffer overflow Considered System Wireless Link, Finite-capacity buffer, Random packet arrivals, Choosing from two alternative power levels. MDP Two alternative power Levels (P1<P2) BPSK modulation ARQ error control Power decisions are made at every time slot. PTs ) Success probability = 1 − Q( t 0 Overflow cost: O ( x, u ) State Transitions Energy Efficiency A key concern in wireless networks, Limited and non renewable power supplies. d i , j (u ) Power Control, Traditional Work: Mitigating effects of interference, Satisfying some QoS constraints, Bambos(96):Maximizing battery lifetime No error control technique is specified. Zorzi(97):Trade-off between instantaneous performance and energy efficiency: ARQ error control, Gilbert-Elliott Channel, Decreasing power in favor of increased retransmissions can be a more energy-efficient. Random Arrivals Previous work generally considered variance of interference and noise. Random arrival of data is not considered generally. In a limited buffer system with random arrivals An arrival burst can occur when queue is long. Buffer overflow can be avoided by increasing power. Changing Arrival rate ÒPu , u 1,2 Ó Ô0, u 0 Single Stage Cost: g ( x, u ) c1 E ( x, u ) c2O ( x, u ), x ´X , u ´U Dynamic Programming Equations: V 0 ( x) min {g ( x, u )} u´U ( x ) V n 1 ( x) min {g ( x, u ) HË d x , x ' (u )V n ( x)} u´U ( x ) x' Optimality of Threshold Policy Changing Discount Factor For 0<x<K optimal policy chooses P2 if: c1 ( P2 P1 ) }H[z( 2 1 )(V n ( x 1) V n ( x)) (1 z)( 2 1 )(V n ( x) V n (x 1))] Minimizing average power subject to average success rate constraint. Interference has a known distribution. E ( x, u ) i 0, j 0 1 z, Ò Õ z i 0, j 1 , Õ Õz K !i !0, j i (1 z)(1 u ), u Õ z(1 u ), K !i !0, j i 1 Õ Ó K !i !0, j i (1 z) u 1 Õ Õ i K, j i 1 u , Õ u , i K, j i 1 Õ Õ 0 | j i |!1 Ô SINR,BER constraints (Foschini(93),Yates(95)) Energy Efficiency Ò0, x !K Ó Ôz, x K Energy Expenditure cost: Optimal policy is of threshold type. Motivation Changing Power Levels VH E x ËHt g ( xt , ut ) x ´X , u ´{0,1,2} N0 Result Computational Results State, xt: Buffer occupancy in the tth slot (xtєX={0,1,…,K}) Control: ut: Power control decision (ut є {0,1,2}) ut is allowed only when xt=0. Find the policy that minimizes the discounted cost: H[z( 2 1 )hxn (1 z)( 2 1 )hxn1 ] (1) Here h V ( x 1) V ( x) Theorem: For P1<P2 the below inequalities hold in every iteration n: n x n n For x greater than or equal to some xn: hxn≤hx+1n (2) For x<xn: c1(P2-P1)≥α(μ1- μ2)(λhxn+(1- λ) hx-1n) (3) For all x: 0≤hxn≤c2 (4) (2) and (3) are useful for proving that the right hand side of inequality (1) crosses left side at most at one point. Practical Determination of Thresholds Let 0≤t≤K be the threshold value. Queue state probabilities π=[π1 π2,…, πK] can be found by solving the Markov chain below: Find t* minimizing the cost below: d1,2(1) d2,3(1) dt-2,t-1(1) dt-1,t(1) dK-2,K-1(2) dK-1,K(2) t 1 t* K arg min{0t g (0,0) Ëxt g ( x,1) Ëxt g ( x,2)} t x 1 x t X=1 X=2 d2,1(1) … X=t-1 X=t d3,2(1)dt-1,t-2(1) dt,t-1(2) … X=K-1 dK-1,K-2(2) X=K dK,K-1(2) Results The optimal power control policy of threshold type. Optimal threshold Increases as P1 approaches to P2, Decreases as the arrival rate increases, Decreases as the discount factor α increases.