In a number of scientific disciplines, researchers are able to observe multiple processes simultaneously, such as stock prices, ecological networks, or neuronal activity, but do not have robust methods to identify how the processes causally influence each other. This talk will examine some recent work in estimating directed information, a quantity shown to characterize statistically causal influences between stochastic processes. A provably good parametric estimation scheme will be shown, and results from an analysis on communication patterns of simultaneously recorded neurons in a primate study will be discussed.
Wireless technologies, such as 802.11, that are used in ad hoc networks provide for multiple non-overlapping channels. In the last decade, many authors have showed the advantages of using multiple channels in ad hoc networks; specifically the available network capacity can be increased. New protocols for routing and MAC layer have been designed to exploit these advantages. Many of these protocols rely on the channel switching capability of the wireless radios to ensure network connectivity. In these kinds of protocols, when a node needs to perform broadcasting, it sends the packet over all the channels available, incurring a switching delay between transmissions on different channels and also introducing to the network a huge and unnecessary overhead. In this work we propose a simple probabilistic approach to perform broadcasting on multiple channels multiple interfaces wireless networks. Even though with a probabilistic approach, there exists some possibility that a broadcast packet is not received by all the nodes in the network, simulation results show us that our approach is sufficient to perform other tasks as neighbor discovering and route request propagation, whilst reducing considerably the overhead required by the simple solution of flooding over all the channels.
We consider the construction of tests for universal hypothesis testing problems, in which the alternate hypothesis is poorly modeled and the observation space is large, motivated by the application in anomaly detection. We investigate a feature-based technique called mismatched universal test for this purpose. Its finite-observation performance can be much better than the (optimal) Hoeffding test, and good performance depends crucially on the choice of features. We propose a feature extraction technique with performance guarantees. This talk is an overview of our research on the mismatched universal test and the feature extraction technique.
We address the problem of finding the minimum decomposition of a permutation in terms of transpositions with non-uniform cost. For arbitrary non-negative cost functions, we describe polynomial-time, constant-approximation decomposition algorithms. For metric-path costs, we describe exact polynomial-time decomposition algorithms. Our algorithms represent a combination of Viterbi-type algorithms and graph-search techniques for minimizing the cost of individual transpositions, and dynamic programing algorithms for finding minimum cost decompositions of cycles. The presented algorithms have applications in information theory, bioinformatics, and algebra.
In this talk we study a simple (to formulate) problem, which we refer to it as the ``Consensus in a Restaurant'' problem. Then we show how the solution to this problem results in an alternative characterization for the ergodicity of a doubly stochastic chain. Finally, we show how the previously known sufficient conditions for the ergodicity of doubly stochastic chains can be deduced from our alternative criterion.
Flow separation is a phenomenon in which the flow encounters an adverse pressure gradient along a surface and is unable to stay attached to it. It is observed in different scenarios, like flows over wings, blades, diffusers and S-ducts and leads to undesirable outcomes like loss of lift, increase in drag, stall and decrease in pressure recovery. In particular reference to jet engines, separation of flow in its diffuser or ducting section adversely affects the performance of the engine. Flow control aims at controlling this separation by using active and/ or passive devices. Two of the many techniques employed are pneumatic actuation, in which a continuous or pulsed jet of air is forced into the core flow, and plasma actuation in which an arc discharge is created between two electrodes on the surface of the body or inside cavities to effectively inject momentum and energy into the core flow. These two actuator systems are being explored to control the flow and encourage quicker reattachment or eliminate separation altogether in open-loop and potentially closed-loop form.
In this talk I will provide a brief introduction to ensemble control theory, which is a new and useful way to deal with bounded uncertainty in dynamical systems. To steer one system with an uncertain parameter, we pretend to steer a continuous ensemble of systems, each with a particular value of that parameter. Despite the corresponding jump from a finite-dimensional to an infinite-dimensional configuration space, I will show how to establish controllability results and to derive practical motion planning algorithms based on these results. Hardware experiments will demonstrate the utility of this approach when applied to real robots.
In this talk, we consider a class of N-person nonzero-sum stochastic games with incomplete information. We develop fully distributed reinforcement learning algorithms, which require for each player a minimal amount of information regarding the other players. We propose heterogeneous and hybrid learning mechanisms that allow players to change their learning schemes at each time based on their rationality and preferences. We apply our results to a class of security games in which the attacker and the defender adopt different learning schemes and update their strategies at random times.
A new imaging technology known as depth cameras are expected to reach the consumer market in the near future. This talk will explore the current research of our lab and explore some of the interesting applications involving depth cameras.
Brain-Computer Interfaces (BCIs), are a nascent technology that allow a human to control a system directly through brain signals. BCI research is a highly interdisciplinary endeavor that synthesizes ideas from neuroscience, biophysics, materials science, control theory, information theory, and signal processing. We will survey the current state of BCI research, and describe current efforts to use error signals generated by humans engaged in a reinforcement learning task to control an autonomous agent.
*Students are sorted alphabetically by family name.
Printer-friendly version