Networks Seminar

Networks Seminar, supported by the Centre for Networked Intelligence, is a technical discussion forum in topics including but not limited to computer networks, machine learning, signal processing, and information theory. The seminar series receives audience from faculty and students in the EECS division, RBCCPS, and engineering professionals working in related fields. Network Seminar series is being held since about an year (Sep. 2019).

Title: Development of an Autonomous Micro-robotic Swarm System – An overview

Speaker:  Akshatha Jagadish
Date: 15/09/2020
 

Abstract: Micro-robotics is an emerging field of research where the focus areas are physical actuation, system control, materials, sensor research and so on. It is an interdisciplinary field where researchers from communities such as physics, chemistry, engineering (biotech, mechanical, comp science) have a role to play. A micro-robot is a controllable machine of micron scale with application specific capabilities in addition to generic functions such as motion, sensing and control mechanism. Scaling robotic systems to micro-scale, forces us to focus on physical parameters such as surface tension, adhesion and drag instead of mass and inertia. There has been research in development of actuation mechanisms at micron scale such as magnetically actuated rigid helices, cilia and sperm-mimetic synthetic tails, chemically powered spherical particles and cannons, synthetically engineered bacteria, muscle cells, etc. Parallel research in this field studies the swarm behaviour and control of such micro-robots. In this talk we will focus on the introduction to this field of research, look at some implementations and talk about our work on the study of the effect of external control on active particle behaviour.

Speaker Bio: Akshatha Jagadish is a PhD student at RBCCPS working with Prof. Manoj Varma (CeNSE). She received her B.Tech. degree in Electronics and Communication from PES Institute of Technology, Bangalore in 2015. She worked as Associate Software Engineer for Automotive Functional Safety at Robert Bosch Engineering and Business Solutions, India during 2015-17. Her current research area is the field of micro-robotic system design and control.

Title: Mathematics of Neural Nets
Speaker:  Anirbit Mukherjee
Date: 08/09/2020
 

Abstract: One of the paramount mathematical mysteries of our times is to be able to explain the phenomenon of deep-learning i.e training neural nets. Neural nets can be made to paint while imitating classical art styles or play chess better than any machine or human ever and they seem to be the closest we have ever come to achieving “artificial intelligence”. But trying to reason about these successes quickly lands us into a plethora of extremely challenging mathematical questions – typically about discrete stochastic processes. Some of these questions remain unsolved for even the smallest neural nets! In this talk we will give a brief introduction to neural nets and describe two of the most recent themes of our work in this direction.

Firstly we will explain how under certain structural and mild distributional conditions our iterative algorithms like “Neuro-Tron”, which do not use a gradient oracle can often be proven to train nets using as much time/sample complexity as expected from gradient based methods but in regimes where usual algorithms like (S)GD remain unproven. Our theorems include the particularly challenging regime of non-realizable data. Next we will briefly look at our first-of-its-kind results about sufficient conditions for fast convergence of standard deep-learning algorithms like RMSProp, which use the history of gradients to decide the next step. In the second half of the talk, we will focus on the recent rise of the PAC-Bayesian technology in being able to explain the low risk of certain over-parameterized nets on standardized tests. We will present our recent results in this domain which empirically supersede some of the existing theoretical benchmarks in this field and this we achieve via our new proofs about the key property of noise resilience of nets.

This is joint work with Amitabh Basu (JHU), Ramchandran Muthukumar (JHU), Jiayao Zhang (UPenn), Dan Roy (UToronto, Vector Institute), Pushpendre Rastogi (JHU, Amazon), Soham De (DeepMind, Google), Enayat Ullah (JHU), Jun Yang (UToronto, Vector Institute) and Anup Rao (Adobe).

Speaker Bio: Anirbit Mukherjee finished his Ph.D. in applied mathematics at the Johns Hopkins University advised by Prof. Amitabh Basu. He is soon starting a post-doc at Wharton (UPenn), Statistics with Prof. Weijie Su. He specializes in deep-learning theory and has been awarded 2 fellowships from JHU for this research – the Walter L. Robb Fellowship and the inaugural Mathematical Institute for Data Science Fellowship. Earlier, he was a researcher in Quantum Field Theory, while doing his undergrad in physics at the Chennai Mathematical Institute (CMI) and masters in theoretical physics at the Tata Institute of Fundamental research (TIFR).

Title: Analysis of Network Logs
Speaker:  Dr. Mouli Chandramouli
Date: 01/09/2020
 

Abstract: Analysis of network logs collected from network devices is presented.  The objective is to understand and determine the important network events and infer the possible root causes of those network events.  The volume of network data is very high and often it can be quite challenging to filter out only the key important messages.  We have developed ML / NLP based techniques to extract the underlying statistical templates of the SYSLOG messages, and secondly identify anomalous patterns observed in the SYSLOG events which can be useful to recommend suitable remedial actions. The proposed solution is under evaluation by network operations.

Speaker Bio: Mouli Chandramouli is currently working as a Data Scientist at Cisco Systems, Bangalore in the area of application of Machine Learning algorithms for analytics of Network Telemetry and Network Inference. He is also a Visiting Professor at the RBCCPS, IISc. He received his M. S. and  Ph.D. from University of Arizona, Tuscon, AZ in the area of Stochastic Process and Queueing Theory.  Prior work experience at AT&T Bell Laboratories, Holmdel, NJ,  Bell Communications Research, NJ in the area of network performance modelling and Dynamicsoft, NJ a startup company focussed on VOIP products based on SIP Protocol which was acquired by Cisco Systems.  At Cisco Systems, his work has been is in the areas of MPLS networks, Energy Management for networking devices and distributed embedded network analytics algorithms.

Title: Minimizing Age in a Multihop Wireless Network   
Speaker:  Ashok Krishnan K. S.
Date: 18/08/2020
 

Abstract: Age of Information (AoI) of packets in a network give a sense of the ‘freshness’ of the information. Applications often require packets to be delivered before they age too much. The talk will discuss a scheduling algorithm designed to transmit packets across the network while meeting age requirements of multiple flows, simultaneously. The algorithm uses a packet dropping rule and a slot wise optimization, which can also be implemented in a distributed fashion. It is seen to perform well, and brings the age close to a theoretical lower bound.

Speaker Bio: Ashok Krishnan K.S. recently completed his Ph.D. from the Department of ECE, IISc. His research interests are in the areas of wireless networks, communications and queueing.

Title: Fair Cake Division    
Speaker:  Nidhi Rathi
Date: 11/08/2020
 

Abstract: The theory of Fair Division addresses the fundamental problem of allocating goods among agents with equal entitlements but distinct preferences. Here, the resources can be (1) divisible like water/land, (2) indivisible like courses in universities, property settlements or (3) indivisible resources with money like electronic frequency allocation. In this talk, I will, in particular focus on the classic cake-cutting problem that provides a model for addressing fair and efficient allocation of a divisible, heterogeneous resource (metaphorically, the cake) among agents with distinct preferences. I will present some of the recent results that complements the existential (and non-constructive) guarantees and various hardness results by way of developing efficient (approximation) algorithms for cake division. I will also talk about a recent result that identifies a broad class of cake division instances that essentially admits a polynomial time algorithm to compute fair and efficient allocations.

Speaker Bio: Nidhi Rathi is an Integrated PhD student at the Department of Mathematics, Indian Institute of Science (IISc). She started her research as a PhD scholar under the guidance of Prof. Siddharth Barman (Dept. of Computer Science and Automation, IISc) and Prof. Mrinal K. Ghosh (Department of Mathematics, IISc). She is a recipient of the prestigious IBM PhD fellowship 2020. Her main area of research is Algorithmic Game theory. In particular, she is interested in exploring the computability of equilibria and fair resource allocations under various settings, and hence, developing algorithms with provable fairness guarantees. She was one of the invited speakers in ACM summer school on Algorithmic Game theory held at IIT Gandhinagar in the summer of 2019.

Title: Support Recovery from Linear Sketches.     
Speaker:  Lekshmi Ramesh
Date: 04/08/2020
 

Abstract: In this talk, I will describe the problem of support recovery, where samples sharing a common unknown support are observed through low dimensional projections or ”linear sketches”, and the goal is to recover the common support. This problem has been well-studied in the single sample setting, and it is known that for certain classes of random projections, the projection dimension only needs to scale linearly in the support size and logarithmically in the dimension of the samples for guaranteeing support recovery. For the multiple sample setting, a natural question to ask is if we can make do with a smaller projection dimension per sample at the cost of a larger number of samples, and to characterize the tradeoff arising between support size, projection dimension, and the number of samples. We will see that the nature of this tradeoff differs depending on whether the projection dimension is larger or smaller relative to the support size. I will mention some results from the literature, which has mostly focused on characterizing this tradeoff in the large projection dimension regime, followed by our results in the case when the projection dimension is small. We will see some commonly used algorithms for this problem, and some lower bounds results that are used to show optimality of the algorithms.

Speaker Bio:Lekshmi Ramesh is a PhD student in the Department of ECE, working with Prof. Chandra R. Murthy and Prof. Himanshu Tyagi.

Title: Metastability phenomenon: large deviations in the stationary regime.                                             Speaker: Sarath AY
Date: 28/07/2020

Abstract: A perturbed dynamical system is said to exhibit the metastability phenomenon when it behaves very differently over different time scales. Many networked systems such as load balancing networks, WiFi networks, etc. exhibit such phenomenon when there are multiple stable operating points in the system. One approach to quantify the metastability phenomenon is to study large deviations of this perturbed dynamics in the stationary regime. This talk will provide an overview of various techniques (in the existing literature) to obtain large deviations in the stationary regime from process-level large deviations.   

Speaker Bio: Sarath is a PhD student in ECE department working with Prof. Rajesh Sundaresan.

Title: Latency of cellular communication                                                                                                         Speaker: Prof. Himanshu Tyagi
Date: 14/07/2020

Abstract: Emerging applications of cellular communication such as control over wireless links require low latency communication. LTE provides an approximately 15ms latency link, which is supposed to suffice for roughly 150-350ms end-to-end latency requirements. The emerging 5G standard will push this envelop further to even lower latency. But standards just pose a challenge which the technology strives to address. In this talk, we will see what are the typical latency requirements for various applications, what are the main factors that add up to LTE latency, and how 5G is trying to circumvent these bottlenecks of latency. It is a summary of our learnings over the past year as a part of the 5G project where we have been trying to understand the standards, study the limitations of typical implementations, and identify the main bottlenecks for low latency communication in practice. This talk is based on weekly discussions with SVR Anand, Aditya Gopalan, Parimal Parag, and other brave soldiers of the 5G V2X team at the ECE department. 

Speaker Bio: Himanshu Tyagi is an Assistant Professor at the Electrical Communication Engineering department at the Indian Institute of Science. He is a co-convenor of the Centre for Networked Intelligence. He received the Bachelor of Technology degree in electrical engineering and the Master of Technology degree in communication and information technology, both from IIT Delhi in 2007. He received the Ph.D. degree in Electrical and Computer Engineering from the University of Maryland, College Park. From 2013 to 2014, he was a postdoctoral researcher at the Information Theory and Applications (ITA) Center, UCSD.

Title: CORNET: A Co-Simulation Middleware for Robot Networks                                                                Speaker: Srikrishna Acharya B.
Date: 30/06/2020

Abstract: CORNET is a co-simulation middleware for applications involving multi-robot systems like a network of Unmanned Aerial Vehicle (UAV) systems. Design of such systems requires knowledge of the flight dynamics of UAVs and the communication links connecting UAVs with each other or with the ground control station. Besides, UAV networks are dynamic and distinctive from other ad-hoc networks and require protocols that can adapt to high-mobility, dynamic topology and changing link quality in power constrained resource platforms. Therefore, it is necessary to co-design the UAV path planning algorithms and the communication protocols. The proposed co-simulation framework integrates existing tools to simulate flight dynamics and network related aspects. Gazebo with robot operating system (ROS) is used as a physical system UAV simulator and NS-3 is used as a network simulator, to jointly capture the cyber-physical system (CPS) aspects of the multi- UAV systems. A particular aspect we address is on synchronizing time and position across the two simulation environments, and we provide APIs to allow easy migration of the algorithms to real platforms.

Speaker Bio:  Srikrishna has a Bachelor degree in Electronics from G. Pulla Reddy College of Engineering & Technology, Kurnool and a Master degree in VLSI and Embedded Systems from the Indraprastha Institute of Information Technology, Delhi.  His current research interest are software and simulation frameworks for connected drones.

Title: The IISc-TIFR agent-based city simulator for the study of COVID-19 spread                                   Speaker: Prof. Prahladh Harsha (TIFR)
Date: 23/06/2020

Abstract: Over the last few months, since the spread of the COVID-19 pandemic, several epidemiological models have been proposed to understand the spread of the infection in the population. In this talk,
we will provide an overview of the agent-based models and compare their pros-and-cons against other models. We will argue that agent-based models let one capture detailed interactions at a granular level and can thus be useful in comparing the impact of various non-pharmaceutical interventions against each other. We will then present the IISc-TIFR agent-based city simulator developed for studying the spread of the COVID-19 infection in the cities of Mumbai and Bengaluru under various “unlocking” strategies. We will
conclude with a presentation of the simulator’s estimates for the infection-spread in Mumbai under different lockdown-relaxation scenarios; containment strategies, phased opening of workplaces, gradual
resumption of train services, importance of compliance.

Speaker Bio: Prahladh Harsha is an Associate Professor at the School of Technology and Computer Science (STCS) at the Tata Institute of Fundamental Research (TIFR), Mumbai. He obtained his Bachelors degree from IIT Madras in 1998 and his Ph.D. from MIT in 2004. After MIT, he was a post-doctoral researcher at Microsoft Research, Silicon Valley, a research assistant professor at the Toyota Technological Institute at Chicago, (2005), a visiting scientist at the University of Texas at Austin and at the Technion, Israel Institute of Technology and has been at the faculty at TIFR since Dec 2009.

Title: Bellman meets Shannon: On the Applications of Dynamic Programming in Capacity Computation           Speaker: V. Arvind Rameshwar
Date: 16/06/2020

Abstract: In this talk, we shall go over dynamic programming-based (or DP-based) methods for computing the capacity of finite-state channels (FSCs) or channels with memory, with and without feedback. First, we consider the setting of FSCs without feedback and derive lower bounds on the capacity for the broad class of input-driven channels, where the current channel state is a time-invariant deterministic function of the previous state and the current input. The lower bounds are based on a DP characterization of a bound on the maximum reverse directed information rate. We show that one can explicitly solve the said DP problem, and in the process, obtain useful achievable rates for the runlength-limited input-constrained binary erasure and binary symmetric channels. This is based on joint work with Prof. Navin Kashyap.

We then move on to the setting of a class of FSCs with feedback, where the feedback capacity expression is amenable to formulation as a DP problem. In particular, we shall consider the inter-cell interference (ICI) channel in NAND flash memories and provide numerical evaluations of the feedback capacity. We also discuss an interesting scenario where a simple constrained code achieves the capacity, with and without feedback. The results are from joint work with Aryabhatt M.R. and Prof. Navin Kashyap.

Speaker Bio:  V. Arvind Rameshwar is a PhD student at the Code Design and Analysis Lab, in the Department of ECE, working under the guidance of Prof. Navin Kashyap. A goldmedallist from BITS Pilani, Hyderabad Campus, he graduated with a B.E. (Hons.) in ECE, in 2018. His research interests lie in the information theory of finite-state channels.

Title: Metastability phenomenon: large deviations in the stationary regime                                              Speaker: Sarath A. Y.
Date: 17/03/2020

Abstract: A perturbed dynamical system is said to exhibit the metastability phenomenon when it behaves very differently over different time scales. Many networked systems such as load balancing networks, WiFi networks, etc. exhibit such phenomenon when there are multiple stable operating points in the system. One approach to quantify the metastability phenomenon is to study large deviations of this perturbed dynamics in the stationary regime. This talk will provide an overview of various techniques (in the existing literature) to obtain large deviations in the stationary regime from process-level large deviations.

Speaker Bio:  Sarath is a PhD student in ECE department working with Prof. Rajesh Sundaresan

Title: Visual Search with a Trembling Hand: An Analysis of Odd Arm Identification in Restless Multi-armed Bandits                                                                                                                                                                    Speaker: Karthik P. N
Date: 10/03/2020

Abstract: This work is motivated by a visual search experiment in which a human subject is shown a number of drifting-dots images. The direction of drift in one of these images (the odd image) is different from the common direction of drift in rest of the images. The goal of the human subject is to identify the location of the odd drifting-dots image in the shortest possible time while keeping his probability of decision error small. Our interest is in understanding the relation between (a) the amount of time taken by the human subject to identify the odd image, and (b) the “closeness” of the odd and the non-odd images used in the experiment. It is often the case that the human subjects participating in such visual search experiments tend to sample image locations uniformly at random in an attempt to complete the given task as soon as possible. In this work, we model the above visual search experiment as a problem of odd arm identification in a multi-armed bandit in which (a) each arm yields Markov observations, and (b) the arms are restless. Further, we model the tendency of human subjects to sample image locations randomly as a “trembling hand” for the human subject, and come up with a metric that captures the notion of “closeness” between the odd and the non-odd images. Our results generalize all the previously known results for odd arm identification in multi-armed bandits.

Speaker Bio:   Karthik is a PhD student in the Wireless Information Systems Lab, Department of ECE, working under the supervision of Prof. Rajesh Sundaresan. Prior to joining for PhD, he served as a project assistant in the Signal Processing for Communications Lab of the Dept of ECE, where he worked with Prof. Chandra R. Murthy. Karthik holds a Bachelor’s degree in Electronics and Communications from RV College of Engineering, Bangalore, where he graduated from in 2014.

Title: The four levels of the fixed-point analysis
Speaker: Prof. Rajesh Sundaresan
Date: 03/03/2020

Abstract: The talk will be on the mean-field limit of an interacting particle system, for e.g., a WiFi system. The fixed-point analysis is a useful technique that helps the analyst understand the system’s equilibrium behaviour. One can identify four levels of fixed points that fix (1) the relationship between certain macroscopic observables of the system, (2) the equilibrium distribution over particle states, (3) the evolution of the mean-field over time, and (4) the law associated with the limiting Markovian evolution of a particle. The talk will highlight these four levels and how they are related to each other.

Speaker Bio:   Prof. Rajesh Sundaresan received the B.Tech. degree in electronics and communication from IIT Madras, India, and the M.A. and Ph.D. degrees in electrical engineering from Princeton University, Princeton, NJ, USA, in 1996 and 1999, respectively. From 1999 to 2005, he worked with Qualcomm Inc., where he was involved in the design of communication algorithms for wireless modems. Since 2005, he has been with the Indian Institute of Science, Bengaluru, India, where he is currently a Professor with the Department of Electrical Communication Engineering and an Associate Faculty with the Robert Bosch Centre for Cyber-Physical Systems. His interests include the areas of communication, computation, and control over networks.

Title: Tracking an Autoregressive process with limited communication
Speaker: Rooji Jinan
Date: 11/02/2020

Abstract: Samples from a high-dimensional AR[1] process are quantized and sent over a communication channel of finite capacity. The receiver seeks to form an estimate of the process in real-time. We consider a time-slotted communication model in slow-sampling regime where multiple communication slots occur between two sampling instants. We propose a successive update scheme which uses communication between sampling instants to refine estimates of the latest sample. We study the following question: Is it better to form refined estimates and send them over multiple communication slots, making the receiver wait more for an update, or to be fast but loose and send new information in every communication opportunity? We show that the fast but loose successive update scheme with spherical codes is universally optimal asymptotically for large dimension. However, most practical quantization codes do not meet the ideal performance required for this optimality, and will typically will have a fixed additive error. Interestingly, our analysis shows that in presence of such an error optimal choice is not fast but loose, but a judiciously chosen frequency of updates is needed.

Speaker Bio:   Rooji Jinan is a 3rd year PhD student at Robert Bosch Centre for Cyber Physical Systems, IISc. She is guided by Dr. Parimal Parag, Department of Electrical Communication Engineering, IISc. Her current research area is timely updates for cyber-physical systems. Rooji received her B.Tech. degree in Electronics and Communication and M.Tech. in Communication Engineering and Signal Processing from Calicut University, Kerala.

Title: Stochastic optimization with compressed gradients
Speaker: Prathamesh Mayekar
Date: 28/01/2020

Abstract: We consider stochastic optimization over $\ell_p$ spaces using access to a first-order oracle. In the first part of the talk, we ask: What is the minimum precision required for oracle outputs to retain the unrestricted convergence rates? We characterize this precision for every $p\geq 1$ by deriving information theoretic lower bounds and by providing quantizers that (almost) achieve these lower bounds. In the second part of the talk, we completely characterize the precision-convergence trade-off for the Euclidean case. Interestingly, the quantizer designed for this setting, RATQ, (almost) achieves the rate-distortion bounds universally for the well-studied Gaussian rate-distortion problem. This talked is based on joint work with Himanshu Tyagi.

Speaker Bio:   Prathamesh Mayekar is a fourth year Ph.D. candidate in the Department of Electrical Communication Engineering at the Indian Institute of Science, Bengaluru. He is advised Dr. Himanshu Tyagi. He received his Master’s degree in Industrial Engineering and Operation Research from the Indian Institute of Technology Bombay in 2015 and a Bachelor’s degree in Electronics and Telecom. Engineering from the University of Mumbai in 2013. Broadly, his research interests lie at the intersection of information theory and optimization. He is a recipient of Jack Keil Wolf ISIT Student Paper Award and Wipro PhD fellowship.

Title: Multi-agent Reinforcement Learning and its applications to Smart Grids
Speaker: Raghuram Bharadwaj
Date: 20/01/2020

Abstract: Reinforcement Learning (RL) deals with the algorithms that a single agent can apply to learn its optimal behavior from the environment. However, in the modern world, we encounter many applications where there are multiple agents involved instead of a single agent. It has many advantages over a single agent learning. For example, consider the task of moving a heavy object from one place to another. It may be impossible for a single agent to complete the task within defined time constraints. We can employ multiple agents and successfully finish the task. The key idea here is that the learning task can be shared among the agents. In this talk, I shall discuss algorithms for multi-agent RL under the settings of constrained cooperative stochastic games and two-player zero-sum games. One of the practical applications of multi-agent learning is Smart Grids. Smart Grid is a concept of developing a power grid that can intelligently make use of electricity. In this context, I shall also be discussing a novel stochastic game framework for energy management in microgrids networks and present the advantages of our proposed framework.

Speaker Bio:   Raghuram is a Ph.D. student in the department of CSA under the guidance of Prof. Shalabh Bhatnagar. His research interests include developing convergent algorithms for multi-agent and off-policy learning in the context of Reinforcement Learning, application of multi-agent reinforcement learning algorithms to smart grids. He is currently working on developing dynamic pricing solutions for energy management in microgrid networks.

Title: Sequential addition of coded tasks for straggler mitigation
Speaker:  Ajay Badita
Date: 14/01/2020

Abstract: Given the unpredictable nature of the nodes in distributed computing systems, some of the tasks can be significantly delayed. Such delayed tasks are called stragglers. In order to mitigate stragglers, redundancy in computation is often employed by encoding k tasks to n tasks such that any k of them can help ascertain the completion of the tasks. Two important metrics of interest are service completion time of the k tasks, and server utilization cost which is sum of time each server spends working on the tasks. Even though starting all n jobs at the start (t = 0) leads to lower mean service completion time, it leads to higher mean server utilization cost. We consider a proactive straggler mitigation strategy where n0 <= n tasks are started at t = 0 while the remaining n − n0 tasks are launched when l0 <= min(n0, k) tasks finish. The tasks are halted when k tasks finish. This gives a flexible forking strategy with multiple parameters. We analyze the mean of two performance metrics for the proposed forking strategy when the random task completion time at each server is independent and distributed as a shifted exponential. This talk demonstrates an effective algorithm to find the tradeoff between the two performance metrics mean server utilization cost and mean service completion time so as to choose efficient choice of parameters. This work has been accepted at INFOCOM-2020 conference.

Speaker Bio: Ajay Badita is a PhD student in the department of ECE, IISc-Bengaluru, working under the supervision of Prof. Parimal Parag.

Title: Large deviations for Cox processes and Cox/G/infinity queues, with a biological application
Speaker: Ayalvadi Ganesh, University of Bristol
Date: 10/01/2020

Abstract: We show that a sequence of Cox processes on a Polish space E satisfy a large deviation principle (LDP), provided their directing measures do so on the space of finite measures on E equipped with the weak topology. Next, we consider a sequence of infinite server queue with general iid service times, where the arrivals constitute Cox processes with translation invariant directing measures assumed to satisfy an LDP. We show that the corresponding sequence of queue occupancy measures also satisfy an LDP. These questions were motivated by the problem of describing fluctuations of molecule numbers in biochemical reaction networks within cells. Joint work with Justin Dean and Edward Cran.

Speaker Bio: Ayalvadi Ganesh received his BTech in EE from IIT Madras in 1988, MS and PhD in EE from Cornell University in 1991 and 1995 respectively. His Ph.D. thesis was on the use of large deviation techniques in queueing theory. He was with Edinburgh University, Birkbeck College, London, U.K., and Hewlett-Packards Basic Research Institute in Mathematical Sciences (BRIMS) and Microsoft Research before joining the Mathematics Department of Bristol University. He was also a Fellow of Kings College, Cambridge, from 2000 to 2004. He has published extensively on Queueing Theory and Large Deviations, Bayes’ Asymptotics, Economics of Communication Networks, Peer-to-peer Systems and Algorithms, Random graphs and stochastic processes on graphs, and Computer Viruses and Worms. He is the coauthor, with Neil O’Connell and Damon Wischik, of the Springer Book “Big Queues” published in 2004. His research interests are in the mathematical modelling of communication and computer networks, and in decentralised algorithms for such networks. Specific interests include large deviations and applications to queueing theory and statistics, random graph models and stochastic processes on graphs, and decentralised algorithms for resource allocation in the Internet and in wireless networks.

Title: Acquisition Games with Partial-Asymmetric Information
Speaker: Prof. Veeraruna Kavitha, Centre for Industrial Engineering and Operations Research, IIT Bombay
Date: 17/12/2019

Abstract: We consider an example of stochastic games with partial, asymmetric and  non-classical information. We obtain relevant equilibrium policies using  a new approach which allows managing the belief updates in a structured  manner. Agents have access only to partial information updates, and our
approach is to consider optimal open loop control until the information  update. The agents continuously control the rates of their Poisson  search clocks to acquire the locks, the agent to get all the locks before others would get reward one. However, the agents have no information about the acquisition status of others and will incur a cost proportional to their rate process. We solved the problem for the case with two agents and many locks and conjectured the results for N-agents. We showed that a pair of (partial) state dependent time-threshold policies form a Nash equilibrium. We further obtained good structural properties of the thresholds.

Speaker Bio: Dr. Veeraruna Kavitha is an Assistant Professor at the Centre for Industrial Engineering and Operations Research (IEOR), Indian Institute Technology Bombay, Mumbai, India, since May 2012. Before joining IITB, she was a Principal Research Scientist at Mymo Wireless, Bangalore and SRM Research Institute, Bangalore, India from December 2011 to May 2012. She was a Post Doctoral Fellow at MAESTRO, INRIA and LIA, University Avignon, France from 2008 to 2011 and a Post-Doctoral Fellow at Tata Institute of Fundamental Research, Bangalore, India from 2007 to 2008. She obtained a Ph.D. degree from Indian Institute of Science, Bangalore, India in 2007 and a M.Sc (Engg) in 2002. Her research interests are broadly in Stochastic processes, Performance Analysis, Queuing Theory, Polling systems, Optimal control, Game theory, Stochastic approximation, and Wireless communications.

Title: Completely Uncoupled Algorithms for Network Utility Maximization
Speaker: Ramakrishnan.S
Date: 17/12/2019

Abstract: The recent advances in wireless systems demands addressing the following resource allocation problems, viz channel selection, user association and power control. The solution to these problems should address the following objectives: (i) Network throughput optimality be ensured (ii) Users get a fair share of the network throughput. Also in a heterogeneous network, where multiple radio technologies coexists a distributed solution is preferable.

In this talk, we present two fully distributed algorithms which provide solutions to the above problems with the stated objectives. We assume that the node’s decisions are based only on their past actions and payoffs which is popularly known as completely uncoupled. Prior work in this setup has focused mainly in maximizing the sum-rate. An important attribute to consider in radio resource allocation is fairness among nodes, i.e. every node should get a fair share of the network throughput. Fairness is taken care of by introducing a utility function of the average rate. Our first algorithm, which we call General Network Utility Maximization (G-NUM), maximizes general non-concave utilities. We show that G-NUM induces a perturbed Markov chain (perturbed by ε), whose stochastically stable states are the set of actions that maximize the network utility. Our second algorithm is motivated by adaptive CSMA algorithms based on Gibbs sampling, where we present an approximate sub-gradient algorithm for concave utilities, which we call Concave Network Utility Maximization (C-NUM). C-NUM is considerably faster and requires lesser memory. Our main contribution is the expansion of the achievable rate region, which the prior works incompletely uncoupled setup has ignored to consider. This expansion aids in allocating a fair share of resources to the nodes.

Speaker Bio: Ramakrishnan received the Bachelors degree in electronics and communication engineering from SCSVMV University, Kanchipuram, India, in 2012.

Title: Operation of water distribution networks
Speaker: Prof. Sridharakumar Narasimhan, Dept. of Chemical Engineering, IIT Madras
Date: 10/12/2019

Abstract: Urban water distribution networks (WDNs) are large scale, complex systems with limited instrumentation. The nexus between water and energy reveals that energy production consumes significant quantities of of water while transporting water for end use is a highly energy intensive operation. Hence, it is important to minimize energy consumption while meeting consumer demands at required pressures On the other hand, if the available water is insufficient or inadequate to meet consumer demands at the required pressures, equitable distribution of the available resource is of primary importance.

The system we consider consists of pumps delivering water to different reservoirs  in a network, with each reservoir catering to time varying demand.  Pumps and ON/OFF valves are used as manipulated variables to control the flow and  pressure. The decision variables  are the number of pumps to be turned on and the state of the valves in the network over a given horizon and the objective  is to minimize energy consumption while meeting the time varying demand.  Given the nonlinear nature of the pump operating curve and the hydraulics, this results in a Mixed Integer NonLinear Program (MINLP). We propose to solve by decomposing it into  series of sub-problems that can be solved efficiently.  Application of these ideas to distribution networks reveals potential significant savings in energy or improvement in supply. Experimental results will be shared followed by our field implementations.

Speaker Bio: Sridharakumar Narasimhan obtained his M.Tech(integrated) and PhD in chemical engineering from IIT Maras and Clarkson University, USA in 1998 and 2006 respectively. He is currently Professor at the Department of Chemical Engineering,IIT Madras, India. His background is in process systems engineering and my interests are broadly in optimal experiment and measurement system design, water distribution networks and continuous manufacturing.

Title: An optimization approach to drift detection and clustering in time-series: Application to air quality data in India
Speaker: Dr. Alexandre Reiffers
Date: 26/11/2019

Abstract: Recent developments in low-cost sensors, wireless network communication, and computational tools have paved the way for applications like monitoring with the high spatial and temporal resolution for example in the context of air quality. However, the reduced quality of sensing units necessitates robust drift detection and calibration schemes. The few existing methods are variants of outlier detection algorithms. We presented an optimization-based clustering algorithm that first smooths the data and then performs clustering for drift detection. We present the detection efficiency of the algorithm with a simulated dataset where the proposed algorithm detects sensor failures like random walks, reduced sensitivity and changes in bias.

Speaker Bio: Alexandre Reiffers is a post-doctoral fellow at Robert Bosch Centre for Cyber-Physical Systems. He received the B.Sc. degree in mathematics (2010) from the University of Marseille, the master degree in applied mathematics (2012) from the University of Pierre et Marie CURIE and the Ph.D. degree in computer science (January 2015) from the INRIA (National research institute in computer science and control) and the University of Avignon. His supervisors were Eitan Altman and Yezekael Hayel. From July 2016 to December 2017, Alexandre Reiffers was a researcher at SafranTech where he was working on comparison of maintenance strategies. Most of his research projects concern the application of mathematical tools (game theory, optimization, stochastic process and machine learning) for a better understanding of real-world problems. The different issues that he studies touch topics such as social networks, speech between human and computer, economy and manufacturing.