In recent years, there is a proliferation of technologies requiring timely information. Examples of the technologies range from smart devices (e.g., smartphones) to smart systems (e.g., smart transportation systems). On the one hand, a smartphone would need timely traffic and transportation information for planning the best route. On the other hand, a smart transportation system would need timely information about vehicles’ positions and speeds for planning collision-free transportation.
In those technologies, end devices (e.g., smartphones and vehicles) can be updated by some remote information sources over wired or wireless networks. The information kept at the end devices should be as timely as possible to accomplish their missions (e.g., planning the best route or planning collision-free transportation). Thus, the age of information was recently proposed in  to measure the freshness of the information at the end devices. Interestingly,  claimed that a throughput-optimal design or a delay-optimal design might not achieve the minimum age of information. Thus, guaranteeing the freshness of information at end devices becomes an issue.
In this paper, we consider the scenario where a mobile device is running an application. The application needs some timely information, which can be downloaded through neighboring access points (APs). However, frequent downloading for minimizing the age of information causes the mobile device a critical power issue. To strike a balance between the age of information and the involved power, the mobile device needs to identify an energy-efficient scheduling algorithm for downloading the latest information.
However, the mobile user can move at will. The connectivity between the mobile device and the APs changes over time; in particular, the resulting channel dynamics is even non-stationary because the mobile device might run the application for a short time. The great uncertainty of the channel dynamics poses the major challenge to the scheduling design. In this paper, we tackle the challenge by developing an online energy-efficient scheduling algorithm, requiring no knowledge about the channel dynamics or the time for running the application.
Our main contribution lies in designing and analyzing online scheduling for power-limited mobile devices with unknown movement and unknown time for running the application. The goal is to minimize a total cost including an age cost and a downloading cost. To reach the goal, we leverage primal-dual techniques 
for linear programs. However, the methodology cannot be applied immediately because the age of information usually involves asquare term (see  for example), resulting in a non-linear program. We successfully address the issue by transforming the scheduling problem into an equivalent scheduling problem in a virtual queueing system. With the transformation, an optimal offline scheduling algorithm can be obtained using a linear program. Then, applying the primal-dual techniques to the linear program, we propose a randomized online scheduling algorithm while showing that the worst-case ratio between the expected total cost incurred by the proposed online algorithm and that incurred by an optimal offline algorithm is (asymptotically) .
I-B Related works
The age of information has been analyzed for several queueing models, e.g., [1, 4, 5, 6, 7]; meanwhile, scheduling for minimizing the average age of information has also been explored, e.g., [8, 9, 10, 11]. Moreover, some works investigated the age-energy tradeoff (with or without energy harvesting) in various scenarios, e.g., [12, 3, 13, 14]. Despite many works on the scheduling design or on the age-energy tradeoff, their design and analysis were based on stochastic models with some stationary assumptions but cannot be applied in non-stationary settings. To fill this gap, in this paper we explore the age-energy tradeoff in a non-stationary setting.
Ii System overview
In this section, we start with describing our network model in Section II-A, followed by defining an age model in Section II-B. Then, we formulate a scheduling problem involving the age-energy tradeoff in Section II-C.
Ii-a Network model
Consider a mobile network in Fig. 1, where a mobile device is moving in an area with some access points (APs). The device is running an application that requires timely information. The information at the device can be updated by downloading the latest information through the APs.
Divide time into slots and index them by , where is the total time for running the application. At the beginning of each slot, all the APs can immediately obtain the latest information from an information source, i.e., we neglect the transmission time between the APs and sources (through wired networks as shown in Fig. 1) while focusing on the bottleneck between the APs and the device (through wireless networks). Because of the device’s mobility, the channel between the device and the APs changes over slots. Let be the state indicating the connectivity in slot : if the device can successfully download the latest information through an AP in slot ; if it cannot access any AP in slot . By we define the connectivity pattern of the mobile device, which is arbitrary with potential non-stationary property.
For each slot with , the device can decide whether to download the latest information. Let be the decision of the device in slot , where if the device decides to download, and if it decides not to download. A scheduling algorithm specifies decision for every slot . A scheduling algorithm is called an offline scheduling algorithm if connectivity pattern (along with running time ) is given as a prior. In contrast, a scheduling algorithm is called an online scheduling algorithm if the connectivity pattern is unavailable; instead, it knows the present connectivity state only.
Ii-B Age model
Let be the age of information at the device at the end of slot , i.e., after decision is made. Suppose that, if the device successfully downloads the latest information, then the age of information becomes zero; otherwise, the age of information increases linearly with slots. Then, the dynamics of the age of information is
where the second case means that the age increases by one if the device cannot download the latest information, because either the device accesses no AP or the device decides not to download. Moreover, we assume that initially the device has the latest information, i.e., .
Ii-C Problem formulation
To investigate the tradeoff between the age of information and the power consumption for downloading the latest information, we define an age cost and a downloading cost as follows. We consider the cost associated with the age of information in slot as , i.e., a linear age cost function. Suppose that the device uses a constant power to download the latest information, incurring a constant cost for each downloading. Regarding non-linear age cost functions and dynamic power allocation, please see Section VI.
Given connectivity pattern , we define the total cost under scheduling algorithm by
where the first term is the downloading cost in slot and the second one is the resulting age cost in slot . In this paper, we aim to develop an online scheduling algorithm such that the total cost can be minimized for all possible connectivity patterns .
However, without knowing the connectivity pattern, an online scheduling algorithm is unlikely to achieve the minimum total cost (obtained by an optimal offline scheduling algorithm). We characterize our online algorithm in terms of the competitiveness against an optimal offline algorithm, defined as follows.
For connectivity pattern , let be the minimum total cost for all possible offline scheduling algorithms. Then, an online scheduling algorithm is called -competitive if
for all possible connectivity patterns , where is called the competitive ratio of the online scheduling algorithm.
With a -competitive online scheduling algorithm, the resulting total cost can be guaranteed to be at most times the minimum total cost, regardless of connectivity pattern and running time .
Iii Primal-dual formulation
In this paper, we approach the scheduling problem by leveraging primal-dual techniques  for linear programs. However, the total age in Eq. (2) would yield square terms (see  for example), resulting in a non-linear total cost function. To resolve the non-linearity, in Section III-A we transform the age model into an equivalent virtual queueing system. Then, by exploiting the virtual queueing system, in Section III-B we successfully formulate a linear program for solving the offline scheduling problem, while proposing a primal-dual formulation. The primal-dual formulation will be used later in Section IV for designing and analyzing our online scheduling algorithm.
Iii-a Virtual queuing system
Fig. 2 illustrates the virtual queueing system transformed from the age model. The virtual queueing system consists of a server, a queue, and some packet arrivals, operating in the same discrete-time system as the mobile network. At the beginning of each slot, a packet arrives at the virtual queueing system, i.e., the -th packet arrives in slot . The server is associated with an ON/OFF channel for each slot, where the channel is ON in slot if and it is OFF if . When the channel is ON in a slot, the server can decide whether to flush the queue in that slot: the server decides to flush the queue in slot if and it decides to idle if . If the channel is ON and the server decides to flush the queue, then the queue becomes empty, i.e., the queue size becomes zero; otherwise, the new arriving packet stays at the queue i.e., the queue size increases by one. The queueing dynamics is identical to the age dynamics in Eq. (1); therefore, the queue size at the end of slot (after decision is made) can be described by .
Suppose that holding a packet at the end of a slot incurs a holding cost of one unit. As packets are left in the queue in slot , the cost for holding all the packets in slot is . Moreover, suppose that each flushing takes a flushing cost of units. The goal for the virtual queueing system is to identify a scheduling algorithm for minimizing the total holding cost plus the total flushing cost over slots. Note that the total cost involved in the virtual queueing system is exactly in Eq. (2). In summary, the scheduling problem in the virtual queueing system is equivalent to the original scheduling problem.
With the transformation, we can find that the original online scheduling problem is related to the classical online TCP-ACK problem , in which a node decides whether to acknowledge received packets for balancing an acknowledgement cost and a delay cost. In fact, the online scheduling problem in the virtual queueing system generalizes the online TCP-ACK problem to noisy channels. In Section VI, we will further generalize to non-linear age cost functions and dynamic power allocation.
Iii-B Primal-dual formulation
According to Section III-A, we will focus on the scheduling problem in the virtual queueing system. To cast the problem into a linear program, we introduce another variable to indicate if the -th packet (arriving in slot ) is in the virtual queueing system in slot , where if it is and otherwise. Since a total of packets arrives at the virtual queueing system by slot , the queue size in slot can be expressed as , i.e., counting for all the packets arriving by slot . The total cost in Eq. (2) then becomes
where the first term is the flushing cost in slot and the second one is the holding cost in slot .
We propose an integer program for solving the offline scheduling problem in the virtual queueing system.
In the integer program, the constraint in Eq. (3b) means that the -th packet arriving by slot (i.e., ) either stays at the virtual queueing system in slot (i.e., ) or has been flushed by slot (i.e., for some ).
By relaxing the integral constraint in Eq. (3c) to real numbers, we can obtain the following linear program.
Linear program (primal program):
The relaxation has no integrality gap (similar to the argument for the ski-rental problem ); thus, a solution for in the linear program can minimize the total cost if connectivity pattern is given in advance, i.e., the solution is an optimal offline scheduling algorithm.
Now, we can see advantages of transforming into the virtual queueing system. The transformation along with the auxiliary variable , for all and , can produce a linear objective function in Eq. (4a) and a linear constraint in Eq. (4b). Next, we refer to the linear program as a primal program and express its dual program as follows.
The primal-dual formulation will be employed in the next section for developing an online scheduling algorithm.
Iv Online scheduling algorithm design
We will develop an online scheduling algorithm using the primal and dual programs formulated in the previous section. For this purpose, we first propose a primal-dual learning algorithm in Section IV-A for obtaining a feasible solution to the primal and dual programs. The primal-dual learning algorithm does not know connectivity pattern ; instead, for each slot it obtains only. At the end of slot , the primal-dual learning algorithm returns a feasible solution to the primal and dual programs. Subsequently, we propose a randomized online scheduling algorithm in Section IV-C
by exploiting the solution from the primal-dual learning algorithm. The underlying idea is that the intermediate solution produced by the primal-dual learning algorithm in each slot can be viewed as the probability of flushing in that slot.
Iv-a Primal-dual learning algorithm
We propose the primal-dual learning algorithm in Alg. 1 for obtaining a feasible solution to the primal and dual programs in the online fashion. All the variables are initialized to be zeros in Line 1. In each new slot , Alg. 1 updates the values of all the variables according to the present channel state . If the channel is ON, then Alg. 1 updates the variables in Lines 1 – 1; if it is OFF, then Alg. 1 does in Lines 1 – 1. The updates are performed by iteration from (i.e., the first packet) until (i.e., the -th packet).
First, consider slot with . The value of in Line 1 implies if the -th packet has been flushed by slot . If , then the -th packet has been flushed; thus, no variable needs to be updated. On the contrary, if , then all the associated variables get updated in Lines 1 – 1.
More precisely, if the condition in Line 1 holds, then variable is updated to be the value of in Line 1 to satisfy the primal constraint in Eq. (4b). Moreover, the intuition of updating in Line 1 is that the value of can imply the probability of flushing the queue in slot . The more packets stay at the queue, the higher the flushing probability is, i.e., the higher the value of is. Thus, for those packets potentially staying at the queue in slot (i.e., satisfying the condition in Line 1), Alg. 1 increases the value of according to Line 1. The constant in Line 1 is specified as in Line 1 for satisfying the dual constraint in Eq. (5b) (see Lemma 5 for detail). Regarding , it is updated to be one for maximizing the dual objective function in Eq. (5a).
Second, consider slot with . The value of remains unchanged (i.e., ) because the server has to idle in that slot. The value of , for , is the same as that in the previous slot (see Line 1) since the server can do nothing for those packets arriving prior to slot . However, is set to be one in Line 1 since the -th packet arriving in slot has to stay at the queue. In addition, similar to the case of , Alg. 1 sets to be one.
Iv-B Analysis of Alg. 1
In this section, we demonstrate the feasibility of the solution produced by Alg. 1, while analyzing the resulting primal objective value. Since the value of variable is updated over slots, we use to represent the value of variable at the end (i.e., after update) of slot . Similarly, let and represent the corresponding values at the end of the -th iteration of slot .
In addition, for each -th iteration of slot , we use , for all , to represent the value of at the end of that iteration, while using to represent the value of at the beginning of that iteration. The two notation sets are just employed for simplifying the following proofs; in fact, in the -th iteration of slot , only one variable can be updated, but for all keeps unchanged.
We first establish the primal feasibility in the next lemma.
See Appendix A. ∎
To examine the dual constraint in Eq. (5b), we need the following technical Lemma 4. In that lemma, for each slot we consider an ordered set whose order follows the processing steps in Alg. 1. Namely, the set can be expressed by , where such that , , . We want to emphasize that the set does not capture all the iterations since slot , e.g, it excludes .
For a slot , assume that the -th element in the ordered set is for some and . If is updated to be one at the end of -th iteration of slot , then at the end of that iteration we have
for all and .
See Appendix B. ∎
We next confirm the dual feasibility using the above lemma.
Considering a fixed slot , we will show that the value of in Eq. (5b) is less than or equal to . Note that , for and with , is updated to be one (in Line 1) if the condition in Line 1 holds. Thus, it suffices to show that at most elements in the ordered set can be updated to be one.
for all and , where the last equality is based on as stated in the lemma. Thus, all the elements in after are no longer updated since their conditions in Line 1 fail. ∎
According to the proof of Lemma 5, the computational complexity of Alg. 1 can be reduced by modifying the iteration in Line 1: consider iteration until where is the -th ON slot before slot , i.e., we remove all the iterations before . That is because the iterations before cannot satisfy the condition in Line 1.
The primal objective value computed by Alg. 1 is bounded above by
for all possible connectivity patterns .
Let and be the increment of the primal objective value and that of the dual objective value, respectively, in the -th iteration of slot . Then, we derive and as follows.
If and , then and because all the variables keep unchanged.
If and , then and .
If and , then , and .
The above four cases can conclude that
for all and .
Iv-C Randomized online scheduling algorithm
In this section, we propose a randomized online scheduling algorithm in Alg. 2. The algorithm updates variable in Line 2 in the same way as Alg. 1 does. Moreover, Alg. 2 uses additional variables: the value of in slot is the cumulative value of until (see Line 2); the value of in slot is the cumulative value of until slot (see Line 2). Let and be the corresponding values at the end of slot .
Alg. 2 picks a uniform random number in Line 2. According to Lines 2 - 2, if and there exists a such that , then the server decides to flush the queue in slot , i.e., the device decides to download the latest information in that slot. The intuition of Alg. 2 is that, with the uniform random choice of , the probability of flushing in slot can be derived as in Appendix C.
The next theorem presents the competitive ratio of Alg. 2.
The expected competitive ratio of Alg. 2 is
approaching as tends to infinity.
We remark that if the server flushed the queue in each ON slot with probability of directly, then the resulting expected total cost would be higher than the primal objective value computed by Alg. 1. Look at the second case of the proof of Theorem 8 (see Appendix C). In that case, the -th packet under the scheduling algorithm can stay at the queue in slot with a non-zero probability, yielding a non-zero expected holding cost in slot . Thus, the holding cost for the packet in slot is higher than the term in the primal objective value computed by Alg. 1.
V Numerical studies
We run Alg. 2 for 10,000 slots with for a fixed value of for all . Note that, under the stationary setting, an optimal offline scheduling algorithm for minimizing the long-run average cost is of the threshold-type (see ). Figs. 5 - 5 display the time-average cost for the proposed Alg. 2 and the optimal offline scheduling algorithm. We can observe that the ratio between the average cost of Alg. 2 and that of the optimal offline scheduling algorithm is at most 1.20 (when and ); moreover, the ratio can even achieve 1.0048 (when and ). The average of the ratios for those results in Figs. 5 - 5 is 1.07. In summary, Alg. 2 performs much better than what we analyzed in the worst-case scenario (with the expected competitive ratio of 1.58).
Moreover, we compare Alg. 2 with an online scheduling algorithm: for each slot with , the server flushes the queue when and idles otherwise. The idea is to make a decision in a greedy way such that the resulting cost in the present slot can be minimized. Figs. 5 - 5 also display the time-average cost for the greedy scheduling algorithm. We can observe that the proposed Alg. 2 significantly outperforms the greedy scheduling algorithm, except for the case when and . The exception is because the greedy algorithm happens to take the threshold close to the optimal one.
In this section, we extend the proposed Alg. 2 to non-linear age cost functions and dynamic power control. Without loss of generality, we can focus on modifying the primal-dual learning algorithm in Alg. 1.
Vi-a Non-linear age cost functions
Let be a function of the age of information, indicating the cost incurred by age . Without loss of generality, we can assume that ; if not, we can scale up the value of and the age cost function to make the function values becomes integers. In this context, packets arrive at the virtual queueing system at the beginning of slot . Then, we can modify the iterations in both Lines 1 and 1 of Alg. 1 to the iterations from to the total number of packets until slot . With the modification, all the previous feasibility results and the competitive ratio hold.
Vi-B Dynamic power control
Next, consider the scenario where the device can adjust its transmission power according to the present channel quality. We focus on the linear age cost function. Let be the minimum cost of downloading the latest information in slot if . Suppose that at most power levels can be used and consider with . Then, by replacing the in Line 1 of Alg. 1 with to make the dual program feasible, we can obtain the same feasibility results while achieving the competitive ratio of
approaching again as tends to infinity.
This paper treated a mobile network where a mobile device is running an application. To realize the mission of the application, the device needs to download the latest information through neighboring access points. The paper proposed a randomized energy-efficient scheduling algorithm for the mobile device to decide whether to download for each slot. In particular, the proposed algorithm enjoys the online feature without the knowledge of the channel dynamics or the time for running the application. The proposed scheduling algorithm can asymptotically achieve the expected competitive ratio of
. Interesting extensions of this work include scheduling for multiple information downloads, imperfect channel estimation, and partial information about the connectivity pattern.
Appendix A Proof of Lemma 3
If and , then the value of is zero. Therefore, the primal constraint holds since
Appendix B Proof of Lemma 4
Second, assume that , for some and , is the -th element in . Suppose that is updated to be one at the end of -th iteration of slot and the claim in Eq. (6) holds.
Then, we consider the -th element in the following. Assume that is the -th element in and it is updated to be one at the end of -th iteration of slot . In that iteration, variable is updated according to Line 1 of Alg. 1, and hence we can obtain
where (a) is due to the value of , for all , at beginning of the -th iteration of slot is greater than or equal to that at the end of -th iteration of slot ; (b) is due to . Then, we can establish the claim in Eq. (6) by
where (a) is due to and ; (b) is based on the inequality in Eq. (7); (c) results from the induction hypothesis for .
Appendix C Proof of Theorem 8
The proof needs the following technical lemma.
The server running Alg. 2 flushes the queue in slot with probability of .
be the continuous uniform random variable whose value is between 0 and 1. Then, under Alg.2 the probability of flushing in slot can be derived in the following two cases.
If , then is the only one value of such that can belong to . Then, the probability of flushing in slot is
If , then similar to the first case the probability of flushing in slot is
Then, we complete the proof. ∎
We proceed to compare the expected total cost incurred by Alg. 2 with the primal objective value in Eq. (4a) computed by Alg. 1. First, according to Lemma 10, the expected number of flushing in slot (under Alg. 2) is . Therefore, the expected flushing cost incurred by Alg. 2 in slot is , which is less than or equal to the term in the primal objective value computed by Alg. 1.
If and , then (similar to Lemma 10) the server running Alg. 2 flushes the queue during the period from slot until with probability of . Therefore, under Alg. 2 the expected number of the -th packet left in the virtual queueing system in slot is
where the inequality is because in Alg. 1 the value of , for all , at the beginning of the -th iteration of slot is less than or equal to that at the end of slot ; additionally, it is less than one in this case. Thus, the expected cost incurred by Alg. 2 for holding the -th packet in slot is less than or equal to the term in the primal objective value computed by Alg. 1.
If and , then with the slot in the fourth case of Appendix A; moreover, the expected cost incurred by Alg. 2 for holding the -th packet in slot is the same as that in slot . Since the expected holding cost for the packet in slot is less than or equal to (according to the above three cases), the holding cost for the packet in slot is less than or equal to as well.
-  S. Kaul, R. D. Yates, and M. Gruteser, “Real-Time Status: How Often Should One Update?” Proc of IEEE INFOCOM, pp. 2731–2735, 2012.
-  N. Buchbinder and J. Naor, “The Design of Competitive Online Algorithms via a Primal–Dual Approach,” Foundations and Trends in Theoretical Computer Science, vol. 3, no. 2-3, pp. 93–263, 2009.
-  A. Arafa and S. Ulukus, “Age Minimization in Energy Harvesting Communications: Energy-Controlled Delays,” Proc. of IEEE ACSSC, pp. 1801–1805, 2017.
-  M. Costa, M. Codreanu, and A. Ephremides, “On the Age of Information in Status Update Systems with Packet Management,” IEEE Trans. Inf. Theory, vol. 62, no. 4, pp. 1897–1910, 2016.
-  A. M. Bedewy, Y. Sun, and N. B. Shroff, “Age-Optimal Information Updates in Multihop Networks,” Proc. of IEEE ISIT, pp. 576–580, 2017.
-  R. D. Yates, “Status Updates through Networks of Parallel Servers,” Proc. of IEEE ISIT, pp. 2281–2285, 2018.
-  E. Najm, R. Nasser, and E. Telatar, “Content Based Status Updates,” Proc. of IEEE ISIT, pp. 2266–2270, 2018.
-  Q. He, D. Yuan, and A. Ephremides, “Optimal Link Scheduling for Age Minimization in Wireless Systems,” IEEE Trans. Inf. Theory, vol. 64, no. 7, pp. 5381–5394, 2018.
-  I. Kadota, A. Sinha, and E. Modiano, “Optimizing Age of Information in Wireless Networks with Throughput Constraints,” Proc. of IEEE INFOCOM, pp. 1844–1852, 2018.
-  Y.-P. Hsu, “Age of Information: Whittle Index for Scheduling Stochastic Arrivals,” Proc. of IEEE ISIT, pp. 2634–2638, 2018.
-  R. Talak, S. Karaman, and E. Modiano, “Optimizing Information Freshness in Wireless Networks under General Interference Constraints,” Proc. of ACM MobiHoc, pp. 61–70, 2018.
-  B. T. Bacinoglu, E. T. Ceran, and E. Uysal-Biyikoglu, “Age of Information under Energy Replenishment Constraints,” Proc. of ITA, pp. 25–31, 2015.
-  S. Feng and J. Yang, “Minimizing Age of Information for an Energy Harvesting Source with Updating Failures,” Proc. of IEEE ISIT, pp. 2431–2435, 2018.
-  S. Nath, J. Wu, and J. Yang, “Optimum Energy Efficiency and Age-of-Information Tradeoff in Multicast Scheduling,” Proc. of IEEE ICC, pp. 1–6, 2018.