Online Algorithms for Computer Networks and Cyber-Physical Systems under Sequential Uncertainty

For both computer networks and cyber-physical systems, the problem inputs are often time-varying, and the decisions much be constantly updated to match the changing inputs. However, making changes to decisions usually incur significant additional cost. For example, changing the placement and sizing of virtualized functions in network function virtualization (NFV) can disrupt the processing of the network traffic. Similarly, there are limits on how fast the power output of large electricity generators can change. Thus, one often faces a challenging tradeoff between a decision that is best for the current input and a decision that can work well for future inputs. These problems thus call for the careful design of ``online algorithms,'' which make decisions based only on currently available inputs, but can also do well under future uncertainty. Our group has made advances for online algorithms in several thrusts.

Competitive Online Algorithms for Micro-Grids

We have studied competitive online algorithms in a micro-grid setting for managing uncertainty both in the flexible demand and in the renewable supply. There are two key novelties. First, note that online algorithms in the typical CS literature often assume absolutely no information about the future. As a result, although they may be shown to attain the smallest possible worst-case competitive ratio compared to the offline solution, in power systems such competitive results are often quite pessimistic because they fail to exploit any partial future information. In contrast, we proposed new models to capture partial future information through (possibly inaccurate) forecasts, and we developed a powerful computational approach that can produce online algorithms with greatly-improved competitive ratios. Second, a common dilemma for online algorithm design is that online algorithms with good competitive ratios may exhibit poor average-case performance. we developed a novel ``Algorithm robustification'' procedure that can take any online algorithm with reasonable average-case performance, and convert it into one with both the optimal competitive ratio and good average-case performance.

Selected publications:

  • Shizhen Zhao, Xiaojun Lin and Minghua Chen, ``Robust Online Algorithms for Peak-Minimizing EV Charging under Multi-Stage Uncertainty,'' IEEE Transactions on Automatic Control, vol. 62, no. 11, pp. 5739-5754, Nov. 2017 [pdf] .
  • L. Lu, J. Tu, C.-K. Chau, M. Chen and X. Lin, ``Online Energy Generation Scheduling for Microgrids with Intermittent Energy Sources and Co-Generation,'' in ACM SIGMETRICS, Pittsburgh, PA, June 2013. [pdf] .
  • Competitive Online Convex Optimization with Switching Costs and Ramp Constraints

    Online convex optimization (OCO) problem has many applications in the context of networking, cloud or edge computing, cyber-physical systems, and machine learning. However, efficient algorithms in many important settings are still unknown. Our work considers the following practical settings. First, we are the first to consider OCO problems with both switching costs and ramp constraints, and develop solutions that resolve the potential infeasibility. Second, we consider the setting with switching cost and look-ahead, and develop the first algorithm whose competitive ratio decreases with the look-ahead window size when the switching-cost coefficient is small, and remains bounded when the switching-cost coefficient is large. We have applied these algorithms to network function virtualization (NFV) and cloud computing.

    Selected publications:

  • M. Shi, X. Lin, S. Fahmy, D.-H. Shin, ``Competitive Online Convex Optimization with Switching Costs and Ramp Constraints," in IEEE INFOCOM, Honolulu, HI, April 2018 [pdf] .
  • M. Shi, X. Lin and L. Jiao, ``Combining Regularization with Look-Ahead for Competitive Online Convex Optimization,'' in IEEE INFOCOM, May 2021 [pdf] .
  • Robust and Efficient Control at Grid Level

    We have also studied robust and efficient control at the Independent System Operator (ISO) and the utility-company levels. Here, the stake is even higher because the power grid in an entire area may be jeopardized if any of the physical constraints in the system are not met. Our work below has jointly considered both the reliability assessment commitment (RAC) and the real-time dispatch problems for an ISO. We developed ``maximally robust algorithms'' that can provably ensure grid safety, whenever there exists any other algorithm that can ensure grid safety under a given level of future uncertainty. Our on-going work will study how to attain similar types of robustness guarantees under more sophisticated settings with energy storage, demand response, and market/incentive mechanisms.

    Selected publications: