Abstracts Track 2022

Area 1 - Methodologies and Technologies

Nr:	10
Title:	Performance Evaluation for Remote Estimator of a Binary Markov Source
Authors:	Yutae Lee
Abstract:	In several emerging networks and applications ranging from cyber-physical systems to timely data analysis, there has been a surge of interest in the freshness of information for real-time remote monitoring problems. When considering the freshness of information, traditional network performance metrics such as throughput and delay are not always effective. The Age of Information(AoI) metric has become a major research topic on real-time monitoring applications. However, recognizing the shortcomings of AoI, researchers have proposed several new metric of information freshness: for example, Age of Synchronization(AoS), Value of Information of Update(VoIU), Urgency of Information(UoI), Age of Incorrect Information(AoII), and Freshness Ratio of Information(FRoI). In the context of monitoring a binary information source over a delay system with feedback, we consider four different performance metrics: real-time error, AoI, AoII, and FRoI. For each performance metric, we model the delay system as a discrete-time Markov chain and evaluate the performance via stochastic analysis. Extensive numerical results are also provided to compare the four performance metrics under various system parameters.

Nr:	11
Title:	An Evolutionary Computation Method for Initial Step Length Selection to Improve Run Time of a Line Search Optimization Algorithm
Authors:	Prashant Prakash Deshpande
Abstract:	Line search algorithms typically employ fairly simplistic strategies for initial step length selection that is subsequently improved upon such that it satisfies Wolfe conditions. We think evolutionary computation will help especially if the user of the algorithm is naïve about the mathematical properties of the function and as a result chooses the seed solution for the algorithm at a fair distance from the optima. We think this is not all that a bad assumption to have in software development research especially in the applied sciences. We report improved strategy performance for the Booth function. Literature Review An unconstrained optimization algorithm popularly uses the iterative Line Search algorithm which sets a starting point for the step length and subsequently computes one that satisfies Wolfe conditions. A successful program converges with a pre-fixed tolerance. A popular strategy in Nocedal and Wright (2006) hereafter referred to as NW uses steepest descent, is easy to implement and is fairly intuitive: InitialStepLength(iteration) = WolfeStepLength(iteration-1)Gradient(iteration-1)UnitDescent(iteration-1)/ Gradient(iteration)*UnitDescent(iteration). This strategy is hereafter referred to as S1. Another algorithm in NW hereafter referred to as A1 chooses 0 as the starting step length for each iteration. We use a combination of both the strategies to improve run time. Our Proposed Algorithm Strategy Only for the first iteration we use A1’s initial step length. For subsequent iterations we use a modified version of S1. We bracket the computation of the step length within a bracket and draw values from this bracket that will satisfy Wolfe conditions. For the first iteration, we begin with an initial step length of zero, compute the step length that satisfies Wolfe conditions and modify it using a seed multiplier to estimate an initial step length for the next iteration. Based on the step length computed in the next iteration we revise the multiplier either upwards or downwards using evolutionary computation. Results and Conclusion We use the 2-dimensional Booth function to test our algorithm’s performance. The minimizer of the function is (1,3) and f(1,3) = 0: f(x,y) = (x+2y-7)^2+(2x+y-5)^2. A cursory look at the function makes it evident that the function reaches its minimum value when the expressions within the bracket are set to 0 and solved simultaneously. We chose this function because users typically treat the optimization algorithm as a black box that pops out a solution and even such cursory observations are routinely not made. Our algorithm converges in 18 iterations and improves upon A1 which converges in 23 iterations but we report a higher error of 0.58% vs A1’s 0.34%. The S1 algorithm doesn’t converge even after 50 iterations. We choose the seed solution at (7,9) to demonstrate the power of our algorithm. Going forward, we will look to improve run time further using advanced numerical linear algebra as we implement the algorithm in a professional software. References Global Optimization Test Problems. Retrieved June 2013, from http://www-optima.amp.i.kyoto-u.ac.jp/member/student/hedar/Hedar_files/TestGO.htm. Nocedal, J and Wright, S.J. 2006. Numerical Optimization. Springer. Wolfe Conditions. Retrieved 2021, from https://pages.mtu.edu/~msgocken/ma5630spring2003/lectures/lines/lines/node3.html.

Nr:	9
Title:	Introduction to Freshness Ratio of Information
Authors:	Yutae Lee
Abstract:	Age of Information (AoI) measures how fresh information is. The AoI refers to the amount of time that has elapsed since the generation of the most recently successfully received message. In this abstract, we explain some conventional AoI metrics and introduce anther metric, Freshness Ratio of Information (FRoI). The metrics related to AoI include average AoI and peak AoI. Average AoI means the time average of AoI and peak AoI means the peak value of AoI. The peak AoI corresponds to the AoI immediately before the information is updated. Consider a case where the freshness of information needs to be kept below a threshold. Average AoI alone cannot determine how much AoI is below the threshold. Peak AoI is sometimes inefficient because it only considers the peak values and not the time interval between peaks. The FRoI is defined to be the fraction of time the age does not exceed a predefined freshness threshold. We are interested in high FRoI in order to maintain fresh information. The FRoI can be utilized in time-critical applications where we need to apply a threshold restriction on AoI.