IDSS Distinguished Seminars

Innovations for the 21st Century Grid

Cryptography rearranges power: it configures who can dowhat, from what. This makes cryptography an inherently political tool, and it confers on the field an intrinsically moral dimension. The Snowden revelations motivate a reassessment of the political and moral positioning of cryptography. They lead one to ask if our inability to effectively address mass surveillance constitutes a failure of our field. I believe that it does. I call for a community-wide effort to develop more effective means to resist mass surveillance. I plead for a reinvention of our disciplinary
culture to attend not only to puzzles and math, but, also, to the societal
implications of our work.

With the emergence of smartphone based sensing for mobility as the main paradigm for sensing in the last decade, radically new information sets have become available for the driving public. This information enables commuters to make repeated decisions on a daily basis based on anticipated state of the network. This repeated decision-making process creates interesting patterns for the transportation network, in which users might (or might not) reach an equilibrium, depending on the information at their disposal (for example knowing or not what other users of the network are experiencing or doing). The present talk starts with a brief presentation of the state of the art in traffic monitoring, leading to a new results in routing games. Routing games offer a simple yet powerful model of congestion in traffic networks, both in transportation and communication systems. The congestion in such systems is affected by the combined decision of the agents (drivers or routers), so modeling the decision process of the agents is important, not only to estimate and predict the behavior of the system, but also to be able to control it. This decision process is often called learning, as agents “learn” information about the system or about the other agents. We propose and study different models of learning with the following requirement: the joint learning dynamics should converge asymptotically to the Nash equilibrium of the game. In particular, we focus on two important properties: Is the model robust to stochastic perturbations (such as measurement noise)? And does the model allow heterogeneous learning (different agents may follow different learning strategies)? We study these questions using tools from online learning theory and stochastic approximation theory. We then present experimental results obtained with an online gaming application in which distributed players can play the routing game: they connect to the web app and participate in the game, by iteratively making decisions about their routes and observing outcomes. We show preliminary results from data collected from the application. In particular, we propose and solve a model estimation problem to estimate the learning dynamics of the players, and compare the predictions of the model to the actual behavior of the players, and discuss extensions and open questions.

This talk will aim to accessibly describe progress on a theory of network architecture relevant to neuroscience, biology, medicine, and technology, particularly SDN/NFV and cyberphysical systems. Key ideas are motivated by familiar examples from neuroscience, including live demos using audience brains, and compared with examples from technology and biology. Background material and additional details are in online videos (accessible from website cds.caltech.edu/~doyle) for which this talk can be viewed as a short trailer. More specifically, my research is aimed at developing a more “unified” theory for complex networks motivated by and drawing lessons from neuroscience[4], cell biology[3], medical physiology[9], technology (internet, smartgrid, sustainable infrastructure)[1][8], and multiscale physics [2],[5],[6]. This theory involves several elements: hard limits, tradeoffs, and constraints on achievable robust, efficient performance ( “laws”), the organizing principles that succeed or fail in achieving them (“architectures” and protocols), the resulting high variability data and “robust yet fragile” behavior observed in real systems and case studies (behavior, data, statistics), the processes by which systems adapt and evolve (variation, selection, design), and their unavoidable fragilities (hijacking, parasites, predation, zombies). A final crucial element is scalable algorithms to allow study and design of complex networks using this theory. We will leverage a series of case studies with live demos from neuroscience, particularly the necessity of layered architectures due to speed accuracy tradeoffs (SAT, e.g. Fitts’ “law”) in vision, cognition, and sensorimotor control. The online videos compare similar laws and architectures from medicine, cell biology and modern computer and networking technology. Zombies emerge throughout as a ubiquitous, strangely popular, and annoying system fragility, particularly in the form of zombie science. In addition to the videos, papers [1] and [4] (and references therein) are the most accessible and broad introduction while the other papers give more domain specific details, most importantly [3] and [9] for biology and medicine. For math details a good place to start is Nikolai Matni’s website (cds.caltech.edu/~nmatni/) or his EECS/IDSS talk on March 9. I’ll be at MIT for several days, and I’m hoping to discuss the implications of this research direction for social systems with IDSS and other researchers. Selected recent references: [1] Alderson DL, Doyle JC (2010) Contrasting views of complexity and their implications for network-centric infrastructures. IEEE Trans Systems Man Cybernetics—Part A: Syst Humans 40:839-852. [2] Sandberg H, Delvenne JC, Doyle JC. On Lossless Approximations, the Fluctuation-Dissipation Theorem, and Limitations of Measurements, IEEE Trans Auto Control, Feb 2011 [3] Chandra F, Buzi G, Doyle JC (2011) Glycolytic oscillations and limits on robust efficiency. Science, Vol 333, pp 187-192. [4] Doyle JC, Csete ME(2011) Architecture, Constraints, and Behavior, P Natl Acad Sci USA, vol. 108, Sup 3 15624-15630 [5] Gayme DF, McKeon BJ, Bamieh B, Papachristodoulou P, Doyle JC (2011) Amplification and Nonlinear Mechanisms in Plane Couette Flow, Physics of Fluids, V23, Issue 6, 065108 [6] Page, M. T., D. Alderson, and J. Doyle (2011), The magnitude distribution of earthquakes near Southern California faults, J. Geophys. Res., 116, B12309, doi:10.1029/2010JB007933. [7] Namas R, Zamora R, An, G, Doyle, J et al, (2012) Sepsis: Something old, something new, and a systems view, Journal Of Critical Care Volume: 27 Issue: 3 [8] Chen, L; Ho, T; Chiang, M, Low S; Doyle J,(2012) Congestion Control for Multicast Flows With Network Coding, IEEE Trans On Information Theory Volume: 58 Issue: 9 Pages: 5908-5921 [9] Li, Cruz, Chien, Sojoudi, Recht, Stone, Csete, Bahmiller, Doyle (2014) Robust efficiency and actuator saturation explain healthy heart rate control and variability, P Natl Acad Sci USA 2014 111 (33) E3476-E3485

Twenty years ago, randomized controlled trials testing social policies were essentially unheard of in developing countries, although there were prominent examples in developed economies. Today their number, scale and scope is much greater than could probably have been imagined. This talk will take stock of the role that randomized controlled trials have played to date, and can play in the future, in guiding policy. We will try to assess both successes and tribulations, challenges and promises.

In the last few decades, we have learned to be careful about causation, and have developed powerful tools for making causal inferences from data. Applying these tools has generated both policy impact and conceptual insights. Prof. Mullainathan will argue that there are a large class of problems where causal inference is largely unnecessary where, instead, prediction is the central challenge. These problems are ideally suited to machine learning and high dimensional data analysis tools. In this talk he will (1) try to delineate the difference between problems that require causation and problems that require prediction; (2) describe results from solving one such prediction problem in detail; (3) highlight the set of new statistical issues these problems raise; and (4) argue that solving these problems can also generate both policy impact and conceptual insights.

The world consists of many interesting things that move: people go to work, home, school, and shop in public transit buses and trains, or in cars and taxis; goods move on these networks and by trucks or by air each day; and food items travel a large distance to meet their eater. Thus, massive movement processes are underway in the world every day and it is critical to ensure their safe, timely and efficient operation. Towards this end, low-cost sensing and acquisition of the movement data is being achieved: from GPS devices, RFID and barcode scanners, to smart commuter cards and smartphones, snapshots of the movement process are becoming available. In this talk, I will present a system for stitching together these snapshots and reconstructing urban mobility at a very fine-grained level. The system, which we call the Space-Time Engine, provides an interactive dashboard and a querying engine for answering questions such as: what is the crowding at a train station? where’re packages held up and how can their delivery be sped up? how can the available supply of transport capacity be better used to address daily demand as well as the demand on exceptional days (such as rallies and severe weather events). I will describe the STE’s capabilities for operational and planning purposes, and as a learning system.

How can we identify the most influential nodes in a network for initiating diffusion? Are people able to easily identify those people in their communities who are best at spreading information, and if so, how? Using theory and recent data, we examine these questions and see how the structure of social networks affects information transmission ranging from gossip to the diffusion of new products. In particular, a model of diffusion is used to define centrality and shown to nest other measures of centrality as extreme special cases. Then it will be shown that by tracking gossip within a network, nodes can easily learn to rank the centrality of other nodes without knowing anything about the network itself. The theoretical predictions are consistent with new field experiments.