seminars

Signal Processing Lab Seminars

AUTOMATIC TRACKING OF MULTIPLE PEDESTRIANS WITH GROUP FORMATION AND OCCLUSIONS Pedro Mendes Jorge Secção de Comunicações e Processamento de Sinais, Instituto Superior de Engenharia de Lisboa 17:00 h, Friday, December 7, 2001 ISR's Meeting Room ABSTRACT
This work addresses the problem of automatic tracking of pedestrians observed by a fixed camera in outdoor scenes. Tracking isolated pedestrians is not a difficult task. The challenge arises when the tracking system has to deal with temporary occlusions and groups of pedestrians. In both cases it is not possible to track each pedestrian during the whole video sequence. However, the system should be able to recognize each pedestrian as soon as he/she becomes visible and isolated from the group. This work presents methods to tackle these difficulties. The proposed system is based on a hierarchical approach which allows the application of the same methods for tracking isolated pedestrians and groups.

ANALYSIS OF MUSICAL EXPRESSIVITY: AN EXPERIMENT ON PIANO IMPROVISATION Filippo Bonini Baraldi Ph.D., Center for Computational Sonology – Department of Electrical Engineering University of Padua 17:00 h, Friday, November 30, 2001 ISR's Meeting Room ABSTRACT
This work focuses on the analysis of the musical communication process through the experimental method. Relations between the emotions to be expressed and the sound features involved are investigated through the analysis of a set of (expressive) performances on a digital piano; perceptual experiments on the perceived emotions are presented as well.
In the musical communication process, emotions, sensations, and feelings, are coded into written notation by the composer. Numerous scientists investigated the musical coded structure searching for generic relations between melodic, harmonic, rhythmic patterns and the emotions that are aroused in the listener. The written code, however, is not the only source of emotional content. In fact, a performance that faithfully reproduces the score, such as one given by a computer, sounds mechanical and inexpressive. The performer, therefore, transforming the musical score into sounds, has to bring back the expressive power of the piece. In order to accomplish this task, the performer must know the context in which the written score was produced, interpreting the ideas of the composer coded into the musical signs. Previous researches attempt to formulate general rules for music performance, i.e. developing computer systems that automatically identify the correct phrasing of the musical score. As a result, “deviations”- from what is prescribed in the score - in timing, dynamics and intonation are considered as being a necessary requisite for a meaningful communication of the composer's ideas. But in addition to the composer's message, the performer wish to transmit his own feelings. Again, deviations in parameters such as tempo, articulation, intensity and timbre are introduced to differentiate various expressive intentions. Relations between different expressive interpretations of a chosen musical piece and the deviations from the nominal values of some acoustical parameters have been found in previous studies. At the end of the music communication channel stands the listener, who receives the sounds from the performer(s) and undergoes an emotional experience. It is been widely demonstrated that experiments on the perceived emotions are important for understanding relations between the performer's intentions, the variables in the sounding music and the listener experience.
In this work we point out some difficulties encountered in those experimental approaches to musical expressivity that are centered on the relation between deviations of musical parameters - source of expressive intentions - and the musical score; i.e. the familiarity of the listener with a given musical piece that can influence the perceived expressivity, or the distance between the cultural contexts of the composer, the performer, and the listener. In order to overtake these kinds of problems we propose a new experimental framework where the figure of the composer is “fused” with the performer's one: musical improvisation. In this way we wish to eliminate the overlap between the composer's and the performer's expressive intentions, realizing a more direct relation between the musical (expressive) message, the sound features, and the listener's experience.
During the presentation the main steps of an experiment on musical improvisation will be outlined, and the main results analysed. Advantages and difficulties of this
approach will be estimated in relation to previous studies on musical expressivity.

MOVING TARGETS IMAGING AND TRAJECTORY ESTIMATION USING ALIASED SYNTHETIC APERTURE RADAR DATA Paulo A. C. Marques Ph.D. Student, Instituto de Telecomunicações, Instituto Superior Técnico - Instituto Superior de Engenharia de Lisboa 17:00 h, Friday, November 16, 2001 ISR's Meeting Room ABSTRACT
This presentation addresses moving targets imaging and trajectory estimation using Synthetic Aperture Radar (SAR). The talk starts with a tutorial review of SAR acquisition and signal processing aspects both for airborne and spaceborne scenarios. The problem of imaging and trajectory estimation of moving objects will then be addressed. A summary of the conventional methodologies to deal with the referred problems and their limitations is made. Finally, I present techniques proposed by our research group to overcome some of the limitations of the classical procedures. These processing schemes make possible the retrieval of the full velocity vector of a moving object using a single SAR sensor and the imaging of fast moving targets using undersampled SAR raw-data. The effectiveness of our proposals is illustrated using simulated and real SAR data from the public MSTAR dataset, which is supplied by the Sensor ATR Division of the Airforce Research Laboratory (USA).

THE BETAI PARAMETERIZED TEMPERATURE CONTROLLER Dejan Milutinovic Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, November 9, 2001 ISR's Meeting Room ABSTRACT
The BetaI predictive controller is a robust controller based on physically meaningful parameters. We propose a particular form of this controller, called BetaI parameterized PI controller. Based on this control law, we develop the controller device nA_Treg96. The built-in estimation procedure makes this device desirable in industrial and laboratory applications. Due to the short tuning time and the easy communication user - control law, the end user is motivated to re-tune control loop.

This controller device is developed as market solution at "Mihajlo Pupin" Institute, Belgrade. The controller is tested on an experimental installation at the Faculty of Agriculture, Belgrade. The presentation will be based on recorded real-time data.

ADAPTIVE SPARSE REGRESSION Mário Figueiredo Professor, Instituto de Telecomunicações, Instituto Superior Técnico 17:00 h, Friday, October 26, 2001 ISR's Meeting Room ABSTRACT
The goal of supervised learning is to infer a functional relation between "input" variables and an "output" variable, based on a set of (maybe noisy) training examples. When the output variable is continuous, we are in the context of regression, whereas in classification problems the output is categorical (a "label"). This functional relation is usually expressed with the help of a set of coefficients (or parameters), and the problem consists in estimating these parameters.

In sparse regression, the goal is to obtain an estimate of the coefficients in which several of them are set exactly to zero. It is well known that sparseness is desirable for several reasons. In this talk, I will present an approach to sparse regression which does not involve any (hyper)parameters that need to be adjusted or estimated.
Experiments with several benchmark data sets show that the proposed approach yields state-of-the-art performance, although it involves no tuning or adjusting of sparseness-controlling parameters.

HIDDEN MARKOV MODELS: THEORY AND APPLICATION TO 2D SHAPE RECOGNITION Manuele Bicego Ph.D. Student, Dipartimento di Informatica, University of Verona, Italy Visiting student, Instituto de Telecomunicações, Instituto Superior Técnico 17:00 h, Friday, October 19, 2001 ISR's Meeting Room ABSTRACT
Hidden Markov Models (HMMs) represent a widespread approach to the modeling of sequences: in the last decade they have been extensively applied in a large number of problems, as speech recognition, handwritten character recognition, DNA and protein modelling, gesture recognition and, more in general, behavior
analysis and synthesis.

The basic theory of Hidden Markov Models is presented in the first part of this talk. The second part addresses their application to the 2D shape modeling: object recognition, shape modeling, and shape classification constitute active research areas in computer vision. Object contours are widely chosen features in order to represent the object, as they are easily estimated from an image and well represent the semantic information also from a perceptual point of view. We show the capabilities of the Hidden Markov Models (HMMs) in modeling shape contours, expressed using a curvature approach, in terms of rotation, occlusion, noise and, preliminarly, tilt and slant projections.

NON-LINEAR TRACKER FOR JOINT ASC/AUV OCEANIC MISSIONS Paulo Oliveira Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, June 22, 2001 ISR's Meeting Room ABSTRACT
In the framework of the ASIMOV project an autonomous underwater vehicle (AUV) is used to collect data that should be sent to a support ship (or to a shore station) via an autonomous surface craft (ASC). Sensors such as a video camera and/or a SONAR are envisioned to be carried by the AUV Infante, to collect scientific data on a pre-specified survey area.

Data exchange between both vehicles should rely on acoustic communications due to the strong attenuation experienced by electromagnetic waves in the water. To achieve high bandwidth acoustic communications, the vertical channel must be used. This constraint leads to the design of joint cooperative missions where the ASC Delfim should be positioned in a vicinity of the vertical position of the AUV with minimal or no exchange of navigation data among both platforms. These requirements lead naturally to the need to implement a tracker on-board the ASC, providing estimates (to the control and mission control systems) on the relative position and velocity of both platforms.

The design of the tracker, modelled as a linear parametrically time-varying (LPV) system. will be described in detail resorting to the use of LMIs. The required sensor packages will be discussed and solutions for two alternative sensor suites will be presented.

ROBUST SHAPE TRACKING IN THE PRESENCE OF CLUTTERED BACKGROUND Jacinto Nascimento Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, June 8, 2001 ISR's Meeting Room ABSTRACT
Many object tracking algorithms are based on low level features detected in the image. Typically, the object shape and position are estimated to fit the observed features. Unfortunately, image analysis methods often produce invalid features (outliers) which do not belong to the object boundary. These features have a strong influence on the shape estimates, leading to meaningless tracking results.
This work proposes a robust tracking algorithm which is able to deal with outliers. The algorithm is based on two key concepts. First, middle level features (strokes) are used instead of low level ones (edge points). Second, two labels (valid/invalid) are considered for each stroke. Since the stroke labels are unknown all labeling sequences are considered and a probability (confidence degree) is assigned to each of them. In this way, all the strokes contribute to track the moving object but with different weights. This allows a robust performance of the tracker in the presence of outliers. Experimental tests are provided to assess the performance of the proposed algorithm in lip tracking and surveillance applications.

PHYSICAL CONSTRAINTS ON THE TIME-FREQUENCY PLANE Paulo Mónica de Oliveira Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, June 1, 2001 ISR's Meeting Room ABSTRACT
Ever since its introduction by Gabor in 1946, the time-frequency plane has presented considerable difficulties. Representing signals in a bidimensional plane with a time and a frequency axis may constitute a proper framework for the spectral analysis of signals with time-varying spectral contents. However, this time-frequency representation of signals is a tool which may originate (and normally will) paradoxical results. The problem is not only of an operational nature, but lies, instead, in the very roots of the concepts involved.
The concept of Instantaneous Frequency, whose very name is so close to a contradiction in terms, is one of them. Its traditional definition, which has now been used for decades, lacks theoretical soundness, and is pretty useless as an operational tool. Another fundamental issue, which is the object of going on discussions, is the issue of spectral resolution. Not knowing what to expect or demand from an analysis tool is, of course, a pervasive handicap. Existing uncertainty relations are evaluated, and its impact (or lack of impact) on time-frequency analysis discussed. New measures of broadness/concentration are presented, and a different, information theoretic approach to the issue of time-frequency resolution is discussed.
Other aspects of time-frequency analysis and procedural subproducts will also be discussed. A very simple and effective procedure to obtain numerical derivatives of sampled sequences is discussed, which avoids many of the shortcomings of the traditional approaches to the issue. A new algorithm for parameter estimation of multicomponent polynomial phase signals is also presented.
Most importantly, the promise of not taking more than 45+15 minutes will be fulfilled.

COMBINED GABOR FILTER SEGMENTATION AND 3D VECTOR QUANTIZATION TO VLBR VIDEO COMPRESSION OF UNDERWATER VIDEO SEQUENCES Jorge Barbosa Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, May 25, 2001 ISR's Meeting Room ABSTRACT
A new very low bit rate video compression technique based on Gabor filters and 3D Vector Quantization is proposed. Oriented Gabor Filters are used to detect and classify image blocks with movement and new 3D Vector Quantization techniques are used to further compress those moving blocks. This approach is originally developed to deal with the problem of the transmission of underwater video sequences of gray level images with hydrothermal vents on a underwater acoustic channel. Due to the physical characteristics of this channel, reliable transmission of digital data using spectrum efficient coherent modulation techniques is severely limited in terms of the achievable transmission rates. Massive data compression is required to achieve acceptable information throughputs due to the huge amount of information of video sequences. Our Gabor filters based tecnhique enable the classification of the image blocks in moving blocks and static blocks based in local energy block determination of dominant motions at each block image and a global threshold corresponding to the global motion in the image sequence. To further increase the compression rate we apply to these moving blocks 3D Vector Quantization techniques. This approach can cope with the transmission rate constraints imposed and provide an acceptable quality of the reconstructed video sequences. When we use this technique to compress standard video test sequences it proves to be efficient.

SPATIAL MODULATION FOR UNDERWATER ACOUSTIC COMMUNICATIONS USING PHASE CONJUGATION João Pedro Gomes Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, May 18, 2001 ISR's Meeting Room ABSTRACT
Spatial diversity using an array of hydrophones is a widely used technique for increasing the robustness of underwater receivers by mitigating the effect of channel variations and reducing the probability of low input SNR due to fading. Motivated by developments in terrestrial wireless communications, multiple transmitters have recently been proposed as an alternative way of increasing the throughput in underwater channels. Unlike wireless radio environments, where physical channels are often modeled as simple time-varying gains, compensating for long-term intersymbol interference caused by very long channel impulse responses is the single most important operation in any underwater communication system. Broadband phase conjugation using a time-reversal mirror provides a way of exploiting the existence of multiple transmitting-receiving transducers in bidirectional communication. The basic technique is relatively simple to implement, and to a large extent avoids making detailed modelling assumptions that are frequently violated in poorly-characterized underwater channels at the frequencies of interest for acoustic telemetry. In this talk, the fundamentals of phase conjugation and the main factors affecting the performance of phase conjugate arrays will be reviewed. More elaborate approaches that are better suited for communications applications will then be described. These include channel estimation techniques for tackling channel variations and the effects of noise, as well as wavefront segmentation for creating parallel channels.

COMPOSITIONAL ABSTRACTIONS OF HYBRID CONTROL SYSTEMS Paulo Tabuada Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, May 11, 2001 ISR's Meeting Room ABSTRACT:
This talk will have to distinct parts:

In the first part I will give an introductory level presentation of hybrid systems, stressing it's importance in the current technological development as well as the exciting scientific challenges created by this new area. I will describe how hierarchies of multiple abstractions can be an effective tool in dealing with the inherent complexity of nowadays applications modeled as hybrid systems.

In the second part I will focus on my work by presenting a very general and abstract framework for control systems, that allows to model, discrete event systems (finite state automata), smooth control systems (differential eqs. based control systems), hybrid control systems as well as some other examples. The concept of abstraction and parallel composition will be formalized in this new control paradigm and I will show that abstractions are compositional by making use of elementary concepts of Category Theory. As an application I will specialize these results for hybrid control systems and show many interesting directions for further research.

DIFFERENT POINTS OF VIEW FOR ASSESSING THE NUMBER OF COMPONENTS IN A MIXTURE MODEL Gilles Celeux Ph.D., INRIA Rhône-Alpes, Grenoble, France 14:00, Wednesday, May 9, 2001 ISR's Meeting Room ABSTRACT:
In this talk, we examine how the difficult problem of choosing a sensible number of components in a mixture model can receive different answers according to the modelling purposes. If the purpose is density estimation the use of the Bayesian Information Criterion (BIC) is advocated. If the purpose is cluster analysis the use of an Integrated Complete Likelihood (ICL) criterion is advocated. Links and differences of both criteria are discussed and are illustrated with numerical experiments.

THE EXTENSION OF THE EULER THEOREM TO n-DIMENSIONAL SPACES Daniele Mortari Ph.D., Universita' degli Studi "La Sapienza" di Roma, Centro di Ricerche Progetto San Marco 17:00, Friday, May 4, 2001 ISR's Meeting Room ABSTRACT:
The general mathematical formulation of the nxn proper orthogonal matrix, that corresponds to a planar rigid rotation in n-Dimensional space, is presented. It is shown that the planar rigid rotation depends on an angle (principal angle) and on a set of (n-2) principal axes. The latter, however, can be more conveniently replaced by only 2 orthogonal directions that identify the plane of rotation. The inverse problem, that is, how to compute these principal rotation parameters from the rotation matrix, is also treated. Then the Euler Theorem is extended to rotations in n-Dimensional spaces by a constructive proof that establishes the intrinsic relationship between the general rotation in n dimensions and a minimum sequence of planar rigid rotations. This fundamental relationship, which introduces a new decomposition for proper orthogonal matrices (those identifying the general rotation), can be expressed either by a product or a sum of the same planar rigid rotation matrices. The similar decomposition, in terms of the skew-symmetric matrices, is also given. Some other properties as well as some numerical examples, are provided.

DESIGN OF AN ACTIVE SONAR FOR MEASURING BUBBLE CLOUDS UNDER BREAKING WAVES Roberta Quant Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, April 27, 2001 ISR's Meeting Room ABSTRACT:
Very broadband sonar signal processing schemes have been developed in recent years for the investigation of ambient noise sources in the ocean. A new development is the application of those signal processing procedures to active sonar systems. It will be shown that good results can be obtained by using a correlation processor. This paper will describe a sonar developed for imaging bubbles below breaking waves from an AUV platform.

SMALL SATELLITES ATTITUDE DETERMINATION METHODS Sónia Marques M.Sc.Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, April 20, 2001 ISR's Meeting Room ABSTRACT:
The determination of an axis spacecraft orientation or attitude has a major role in guidance, navigation and control of an aerospace vehicle, especially for autonomous systems which are less fault tolerant than ground-based systems. This work was motivated by the need to develop a method to estimate both the attitude and the angular velocity of a low orbit small satellite, in particular the Portuguese satellite, PoSAT-1. This class of satellites has a number of characteristics such as mass and volume constraints, the non-linear equations of motion and noisy measurements from the attitude sensors which are not available along the whole orbit of the satellite.
Satellite attitude determination methods usually fall in one of two classes: point-to-point and recursive estimation algorithms. Point-to-point attitude determination is based on the measurements of two or more sensors in a single point in time, while recursive estimation uses information from successive time points, as well as knowledge about the spacecraft attitude dynamics model. In small satellites, only a single attitude sensor is often available, due to cost and space constraints, thus leading to the exploration of recursive estimation based solutions, such as the Kalman filter. In this work, the results of using a point-to-point Singular Value Decomposition (SVD) algorithm are compared to those obtained by an Extended Kalman Filter (EKF), when applied to a simulation of PoSAT-1. Results from the EKF, applied to the small satellite PoSAT-1 real data, is also presented.

SCALING LAWS: A WAVELET CHARACTERIZATION Paulo Gonçalves Ph.D., INRIA Rhône-Alpes, Grenoble, France 15:00, Wednesday, April 18, 2001 ISR's Meeting Room ABSTRACT:
Wavelet tools have been widely and successfully used in the analysis of fractal signals. After recalling some basics on fractals and multifractals, we illustrate on several examples of 1/f-like power spectra, which peculiar properties of the wavelet transforms play a key-role identifying and characterizing scaling properties. In the course of our talk, we will address some open issues, among which the problem of existence of high order moments for a random variable. In this direction we will propose a simple wavelet-based test aimed at certifying reliability of empirical estimates.
To conclude, and if time allows, we will discuss a new “non-degenerated multifractal” process with associated scaling properties.

OBSERVER-BASED CONTROL OF PIECEWISE-AFFINE SYSTEMS Luís Rodrigues Ph.D. Student, Dept. of Aeronautics and Astronautics, Stanford University, USA 17:00, Friday, April 6, 2001 ISR's Meeting Room ABSTRACT:
Piecewise-affine systems represent an important and powerful model class to approximate nonlinear systems. This work presents a new Lyapunov-based synthesis method for both state and output feedback control of piecewise-affine systems. The synthesis method builds on existing stability analysis tools by transforming controller design into an analysis problem for the closed loop system where the controller parameters are unknowns. More specically, the proposed technique formulates the control design as an optimization problem subject to linear constraints and a Bilinear Matrix Inequality. This optimization problem can be solved iteratively (for a local maximum) as a set of two convex optimization problems involving linear matrix inequalities. The latter can be solved numerically very efficiently using currently available software packages.
Key points of this synthesis technique are its generality, the fact that it can be completely automated once the system dynamics are given and also that it can be used to design controllers with different structures depending on the number of constraints that are added. In particular, it is shown that a controller with the structure of a regulator and an observer can be designed so that switching based on state estimates rather than on the output can be performed. It is also shown that many other
desired features can be included in the design, such as boundedness of the control signals and avoidance of sliding modes.
It will be shown that this work enables a completely automated synthesis tool for a piecewise-affine approximation of a class of nonlinear systems.
The design of controllers for the original nonlinear system can then be posed as a robust control problem. The applicability of this design method to systems with multiple equilibrium points will be shown in simulation examples.
A toolbox for MATLAB developed by the author was used to perform the designs whose performance is shown in the simulations.

ON COOPERATION AND AUTONOMY IN MULTIAGENT AND MULTIROBOT SYSTEMS Gerhard Kraetzschmar Professor, University of Ulm, Germany 17:00, Friday, March 30, 2001 ISR's Meeting Room ABSTRACT
In most multiagent systems, the agents are supposed to somehow cooperate, while most mobile robot systems are supposed to be autonomous. Multirobot systems, in particular teams of autonomous mobile robots that cooperate to achieve a common task, such as RoboCup soccer teams, force researchers and designers to re-think the relationship between cooperation and autonomy. The talk will illustrate concepts, problems, and solutions using various examples from own prior and current research, ranging from distributed reason maintenance to service robotics applications.

A MARKOV POINT PROCESS FOR ROAD EXTRACTION IN REMOTELY SENSED IMAGES Josiane Zerubia Ph.D., INRIA Sophia-Antipolis, France 15:00, Wednesday, March 28, 2001 ISR's Meeting Room ABSTRACT
We present a new method to extract roads in remotely sensed images. Our approach is based on stochastic geometry theory and reversible jump Monte Carlo Markov Chains dynamic. We consider that roads consist of a thin network in the image. We make the hypothesis that such a network can be approximated by a network composed of connected line segments. We build a marked point process, which is able to simulate and detect thin networks. The segments have to be connected, in order to form a line-network. Aligned segments are favored whereas superposition is penalized. Those constraints are taken in account by the prior model (Candy model), which is an area-interaction point process. The location of the network and the specifities of a road network in the image are given by a term based on statistical hypothesis tests. In order to avoid local minima, a simulated annealing algorithm, using a reversible jump MCMC dynamic is designed. Results are shown on SPOT, ERS and aerial images.

MULTI - SENSOR NAVIGATION FOR SOCCER ROBOTS Carlos Marques Ms.C. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, March 23, 2001 ISR's Meeting Room ABSTRACT
This seminar presents a method for robot navigation in structured indoors environments, based on the information of multiple sensors. A catadioptric omni-directional vision sensor is used for self-localization, from the extraction of natural geometric landmarks of the environment. It is assumed that the robot moves on flat surfaces and straight lines can be identified in the surrounding environment by the catadioptric system. The guidance control algorithm, which takes the robot to a desired posture, is based on odometry and its periodic reset by the self-localization system, and uses sonar data to avoid and move around obstacles. Results from the application to a real soccer robot moving on a RoboCup soccer field are presented.

EMOTIONS IN AUTONOMOUS ROBOT LEARNING Sandra Gadanho Post-Doc, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, March 16, 2001 ISR's Meeting Room ABSTRACT
The adaptive value of emotions in nature indicates they might also be useful in artificial creatures. The work presented investigates ways in which artificial emotions can improve behaviour in the domain of a simple, but complete, solitary learning agent. For this purpose, a non-symbolic emotion model was designed and integrated in a behaviour-based reinforcement-learning framework. Several roles of emotions were investigated: emotions as reinforcement, emotions influencing perception and emotions determining when to re-evaluate the behaviour selection.
Experiments were carried out in a realistic robot simulator that compared the performance of emotional with non-emotional agents in a survival task that consists of maintaining adequate energy levels in an environment with obstacles and energy sources.

ALIGNMENT-BY-RECONSTRUCTION FOR 3D ULTRASOUND IMAGING João Sanches Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, March 9, 2001 ISR's Meeting Room ABSTRACT
This presentation addresses 3D reconstruction of human organs from a set of ultrasound images, assuming that the probe position and orientation is available. Unfortunately, pose measurement errors produce significant misalignments which degrade the performance of the reconstruction algorithms. In this presentation we propose a method for correcting these distortions by estimating the true position and orientation of the ultrasound probe. This is achieved by a joint computation of the 3D data and probe parameters using a MAP criterion. Some experimental results are shown including some tests using the Monte Carlo method to evaluate the performance of the algorithm.

ALGEBRA AND ALGORITHMS FOR QoS PATH COMPUTATION AND HOP-BY-HOP ROUTING IN THE INTERNET João Luís Sobrinho Professor, Instituto de Telecomunicações, Instituto Superior Técnico 17:00, Friday, March 2, 2001 ISR's Meeting Room ABSTRACT
Prompted by the advent of QoS routing in the Internet, we investigate the properties that path weight functions must have so that hop-by-hop routing is possible and optimal paths can be computed with a generalized Dijsktra's algorithm. For this purpose we define an algebra of weights which contains a binary operation, for the composition of link weights into path weights, and an order relation. Isotonicity is the key property of the algebra. It states that the order relation between the weights of any two paths is preserved if both of them are either prefixed or appended by a common, third, path.
We show that isotonicity is both necessary and sufficient for a generalized Dijkstra's algorithm to yield optimal paths. Likewise, isotonicity is also both necessary and sufficient for hop-by-hop routing. However, without strict isotonicity, hop-by-hop routing based on optimal paths may produce routing loops. They are prevented if every node computes what we call lexicographic-optimal paths. These paths can be computed with an enhanced Dijkstra's algorithm that has the same complexity as the standard one. Our findings are extended to multipath routing as well.
As special cases of the general approach, we conclude that shortest-widest paths can neither be computed with a generalized Dijkstra's algorithm nor can packets be routed hop-by-hop over those paths. In addition, loop-free hop-by-hop routing over widest and widest-shortest paths requires that each node computes lexicographic-optimal paths, in general.

ATTENTION MECHANISMS AND VISUAL TRACKING Alexandre Bernardino Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00, Friday, February 16, 2001 ISR's Meeting Room ABSTRACT
Despite the effort of several decades of research, the performance of artificial vision systems in general environments is still very poor when compared with biological systems. In the beginning it was believed than many difficulties could be solved with the increase of computational power and sensor quality. However, in the early 90's, the computational complexity of visual tasks was addressed, and Tsotsos proved that some visual search problems are intratable (NP-complete). It was realized that different approaches should be explored, and work on the analysis of biological systems began to appear.
In this talk, we present some features of biological visual systems that address the visual tracking problem in more efficient ways than conventional techniques. These features are related to attention mechanisms in the sense that visual and computational resources are allocated specifically to the task at hand. A dynamic focus of attention directs computational resources to predicted target positions (covert attention), increasing the amount of object related data available for motion estimation. Motion is then actively tracked by a moving foveated sensor (overt attention) that allocates visual resources to the center of the observation direction. Attention mechanisms are used together with ocular movements to develop a real-time tracking system based on object appearance.

FEASIBLE FORMATIONS OF MULTI-AGENT SYSTEMS Paulo Tabuada Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, February 9, 2001 ISR's Meeting Room ABSTRACT
Advances in communication and computation have enabled the distributed control of multi-agent systems. This philosophy has resulted in next generation of automated highway systems, coordination of aircraft in future air traffic management systems, as well as formation flying aircraft, satellites, and multiple mobile robots. The control of multiple homogeneous or heterogeneous agents raises fundamental questions regarding the formation control of a group of agents. Multi-agent formations require individual agents to satisfy their kinematics while constantly satisfying inter-agent constraints. In typical leader-follower formations, the leader has the responsibility of guiding the group, while the followers have the responsibility of maintaining the inter-agent formation. Distributing the group control tasks to individual agents must be compatible with the control and sensing capabilities of the individual agents. As the inter-agent dependencies get more complicated, a systematic framework for controlling formations is vital. In this talk, I will propose a framework for formation control of multi-agent systems focusing on the feasibility problem: "Given the kinematics of several agents along with the inter-agent constraints, determine whether there exists agent trajectories that maintain the constrains"; and the formation control abstraction problem: "Given a feasible formation, extract a smaller control system (the abstraction) that maintains formations along its trajectories."

SCENE LEVEL RATE CONTROL ALGORITHM FOR MPEG-4 VIDEO CODING Paulo Nunes Ph.D Student, Instituto de Telecomunicações, Instituto Superior Técnico 17:00 h, Friday, February 2, 2001 ISR's Meeting Room ABSTRACT
Object-based coding approaches, such as the MPEG-4 standard approach, where a video scene is composed by several video objects, require that the rate control is performed by using two levels: the scene rate control and the object rate control. In this context, this paper presents a new scene level and object level rate control algorithm for low delay MPEG-4 video encoding capable of performing bit allocation for the several VOs in the scene, encoded at different VOP rates, and aiming at obtaining a better trade-off among spatial and temporal quality for the overall scene. The proposed approach combines rate-distortion modeling using model adaptation by least squares estimation and adaptive bit allocation to 'shape' the encoded data in order that the overall subjective quality of the encoded scene is maximized.

Keywords: MPEG-4 video coding, object-based bit rate control, rate-distortion modeling

COMBINED PLANT CONTROLLER OPTIMIZATION FOR AUTONOMOUS UNDERWATER VEHICLES Carlos Silvestre Professor, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, January 26, 2001 ISR's Meeting Room ABSTRACT
The talk addresses the following combined plant/controller optimization (PCO) problem: given an AUV with a fixed baseline configuration, determine the optimal size of the vehicle control surfaces so that a weighted average of the power required along a given mission is minimized, subject to open and closed loop control requirements. The solution proposed is rooted in LMI theory and provides efficient PCO algorithms that make it possible to assess the tradeoffs involved in the combined design of plant and feedback controllers for autonomous vehicles.

APPLICATION OF BEAM - FORMING ADAPTIVE ALGORITHMS FOR THE UNIVERSAL MOBILE COMMUNICATIONS SYSTEM João Miguel Gil Ph.D. Student, Instituto de Telecomunicações, Instituto Superior Técnico 17:00 h, Friday, January 19, 2001 ISR's Meeting Room ABSTRACT
It is known that beam-forming, understood as the process to control antenna array weights in order to optimize a certain cost function, is not new in mobile communications. Beam-forming is hereby understood as a process independent of the high resolution, direction estimation algorithms. Adaptive beam-forming solutions applied to mobile communications have also been widely covered and studied by many people. This presentation shows how the application of the main beam-forming algorithms and possible structures relate to the application to Universal Mobile Communications system (UMTS). Besides the algorithm-UMTS relationship, the work presented aims at the application within a Wideband Directional Channel Model (WDCM). Likewise, it is a system-propagation-algorithm oriented presentation. In the first part, the talk will cover the WDCM issue, by itself, presenting several directional models and their major characteristics, aiming at a comparison between these. In the second part, an application of the Conjugate Gradient algorithm to defined, still simplified WDCM and UMTS scenarios, is presented, describing the nature of the problem, the convergence analysis and the resulting conclusions and problems.

ALGEBRAIC ASPECTS OF RECONSTRUCTION OF STRUCTURED SCENES FROM ONE OR MORE VIEWS Etienne Grossman Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, January 12, 2001 ISR's Meeting Room ABSTRACT
We consider the problem of 3D reconstruction when the input data consists of image features -points and lines- localized in one or more images and knowledge of relations that hold between some of these features: alignment, coplanarity, orthogonality or parallelism. It has been shown that this kind of input is sufficient, even when a single image is present, to obtain 3D reconstruction. We give a mathematical proof of this statement and present a new method to solve this problem. A single linear system is built whose solution is the estimated reconstruction. Three other characteristics of the method are that it works with one or more images indifferently; ambiguities in the reconstruction problem are clearly indicated and there is no preferential choice of coordinate system or reference object. We focus on theoretical aspects, trying to provide sufficient mathematical rigor while avoiding excessive detail.

STRUCTURE AND PERFORMANCE OF THE ASIMOV ACOUSTIC COMMUNICATION SYSTEM João Pedro Gomes Ph.D. Student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, January 5,2001 ISR's Meeting Room ABSTRACT
In the framework of project ASIMOV, an autonomous underwater vehicle and an autonomous surface craft cooperate to jointly carry out complex missions at sea. Given the widely-varying speed and reliability requirements of traffic between the two vehicles, two distinct communication links were installed. One of them uses robust non-coherent muodulation, achieving data rates of at most 400 bps. The other is a high-speed coherent link operating at 30 Kbps, which can only be used for near-vertical transmission. This talk will mainly address the issues that are relevant for the design of the high-speed link. In particular, the receiver algorithms must be kept relatively simple to enable real-time operation on a single DSP board with low power consumption, yet they must tolerate moderate intersymbol interference and channel fluctuations. After describing the signal processing and coding structure on both ends of the link, experimental results from the 1999 and 2000 missions in the Azores will be presented.

INTERFEROMETRIC ABSOLUTE PHASE RECONSTRUCTION IN SAR/SAS: A BAYESIAN APPROACH José Bioucas Dias Professor, Instituto de Telecomunicações, Instituto Superior Técnico 17:00 h, Friday, December 15, 2000 ISR's Meeting Room ABSTRACT
Absolute phase estimation/reconstruction from incomplete, noisy, and modulo-2p observations is a necessary step in many imaging applications. Some examples are interferometric synthetic aperture radar (InSAR) and sonar (InSAS), magnetic ressonance, optical interferometry, and diffraction tomography. In this talk I will present a new approach to absolute phase (not simply modulo-2p) estimation from incomplete noisy and modulo-2p observations in interferometric type application. The Bayesian viewpoint is adopted to combine the observation mechanism and the absolute phase prior knowledge; the observation density is 2p-periodic and accounts for the interferometric pair decorrelation and the system noise; the a priori probability of the absolute phase is modeled by a
Compound Gauss Markov random field (CGMRF) tailored to piecewise smooth absolute phase images, thus modeling discontinuities present in the absolute phase. To determine the the absolute phase estimate we propose an iterative scheme aiming at the computation of the maximum a posteriori probability (MAP) estimate. Each iteration embodies a discrete optimization step, implemented by network programming techniques, and an iterative conditional modes (ICM) step. I will present a set of experimental results illustrating the effectiveness of the proposed scheme compared with classical approaches.

DISTRIBUTED SCHEDULING WITH ACTIVE IDLENESS: A KEY TO MULTI - CLASS QUEUING NETWORKS STABILIZATION José Moreira MSc. student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, November 24, 2000 ISR's Meeting Room ABSTRACT:
In this talk we will show that any distributed scheduling policy can be changed in order to ensure the resulting queing network to be stable. The traffic condition has to be observed as well as some mild assumptions on the service time and arrival process distributions.
We will present a Time Window Controler, which is an implementation of the Active Idleness concept. By active idleness we mean the controler's ability to send its associated server into a period of idleness in situations where there are customers to be served. It will be shown why this ability is crucial to stabilize and even improve performance in multi-class, non acyclic, and stochastic queing networks.
Experimental results obtained so far will illustrate how one can achieve stability as well as performance improvement over non-idling policies. A series of reported unstable systems will be shown to be stable under our policies.

BLIND IDENTIFICATION OF MULTIPE-INPUT/MULTIPLE OUTPUT WIRELESS CHANNELS BASED ON SECOND ORDER STATISTICS: THEORY, CLOSED-FORM/ITERATIVE ALGORITHMS AND PERFORMANCE ANALYSIS João Xavier PhD student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, November 17, 2000 Department of EEC, Meeting Room, 5th floor ABSTRACT
1. Problem formulation
Identification of multiple input-multiple output (MIMO) systems is a problem of interest in many applications, e.g., SDMA (space division multiple access) networks for wireless comunications, where the base station observes the mobile sources through an unknown linear convolutive mixing channel. To reject both intersymbol inteference (ISI) and interuser inteference (IUI), the SDMA receiver must identify the unknown MIMO system parameters. In this talk, we consider the problem of blind identification of MIMO systems, based only on second order statistics (SOS), i.e., given the correlation matrices of its output. Current SOS methods only solve the special case of static systems (IUI, but no ISI). For the general case of convolutive systems (IUI and ISI), extra higher order statistics (HOS) methods are required to complete the identification.
2. Contribution
(a) We present a theoretical system identifiability result which asserts that general MIMO systems (IUI and ISI) can be uniquely identified from the SOS of its output, when its inputs satisfy a certain SOS condition (otherwise, the identification may be unfeasible).
(b) We prove that the sufficient input SOS condition in (a) can be induced by pre-filtering the inputs with minimum-phase correlative filters of arbitrary non-zero degree.
(c) The proof of the identifiability theorem in (a) is non-constructive. We develop a closed-form algorithm which attains the identifiability bound, thus solving in an integrated manner (no extra algorithm required) the problem of blind identifiaction of general MIMO systems based only on SOS. To account for practical distortions (e.g., carrier phase drifts), an iterative procedure is presented to refine the results of the analytical solution.
(d) The asymptotic mean-square error of the analytical MIMO estimate is obtained in closed-form by using tools from large sample theory (m-dependent sequences, Cramer-Wold device, Slutsky theorems) and validated against numerical experiments. Also, we compare our approach with current alternative solutions, through computer simulations, in typical multipath channels.

PERIODIC DECIMAL FRACTIONS - AN INTRODUCTION Fernando Rebelo Simões Professor, Instituto de Sistemas e Robótica, Instituto Superior Técnico 17:00 h, Friday, November 10, 2000 ISR's Meeting Room ABSTRACT
A simple approach of periodic decimal fractions of the form n/p, n<p, in which p is a prime number is presented. This approach is deliberately made in an experimental and intuitive basis, starting from observations towards formal conclusions, along with some conjectures.
Two main features are emphasized:
- the circular way in which the periods run
- the prime factoring of the generating function 10^n -1, corresponding to a period length n.

UNSUPERVISED LEARNING OF FINITE MIXTURE MODELS Mário Figueiredo Professor, Instituto de Telecomunicações, Instituto Superior Técnico 17:30 h, Friday, November 3, 2000 ISR's Meeting Room ABSTRACT
Finite mixtures are a flexible and powerful probabilistic modeling tool for univariate and multivariate data. The usefulness of mixture models in any area which involves statistical modelling of data (such as pattern recognition, computer vision, signal and image analysis, machine learning) is currently widely acknowledged.
In this talk, I will present a new unsupervised algorithm for learning a finite mixture model from multivariate data. The algorithm is "unsupervised" because: (i) it is capable of selecting the number of components, and, (ii) unlike the standard "expectation-maximization" (EM) algorithm, it does not require careful initialization. The proposed method also avoids another well-known drawback of EM for mixture fitting: the possibility of convergence towards a singular estimate at the boundary of the parameter space.
The novelty of the approach is that it seamlessly integrates estimation and model selection in a single algorithm. The technique can be applied to any type of parametric mixture model for which it is possible to write an EM algorithm; in this talk I will illustrate it using Gaussian mixtures and mixtures of factor analyzers.

ROBUST POINT CORRESPONDENCE BY CONCAVE MINIMIZATION João Maciel PhD student, Instituto de Sistemas e Robótica, Instituto Superior Técnico 15:00 h, Friday, October 27, 2000 ISR's Meeting Room ABSTRACT
We propose a new methodology for reliably solving the correspondence problem between points of two or more images. This is a key step in most problems of Computer Vision and, so far, no general method exists to solve it.
Our methodology is able to handle most of the commonly used assumptions in a unique formulation, independent of the domain of application and type of features. It performs correspondence and outlier rejection in a single step, and achieves global optimality with feasible computation. Feature selection and correspondence are first formulated as an integer optimization problem. This is a blunt formulation, which considers the whole combinatorial space of possible point selections and correspondences. To find its global optimal solution we build a concave objective function and relax the search domain into its convex-hull. The special structure of this extended problem assures its equivalence to the original one, but it can be optimally solved by efficient algorithms that avoid combinatorial search.

IMPORTANCE SAMPLING IN THE ANALYSIS OF DIGITAL COMMUNICATION SYSTEMS: A CASE STUDY Francisco Sena da Silva PhD Student, Instituto de Telecomunicações, Instituto Superior Técnico 15:00 h, Friday, October 20, 2000 ISR's Meeting Room ABSTRACT
Performance analysis in digital communications using Monte Carlo may imply long simulation times especially when low error rates are being estimated. Importance Sampling is a variance reduction technique that allows considerable simulation time gains when compared with conventional Monte Carlo for the same required precision. When designing Importance Sampling a careful modification of the simulation parameters must be done in order to achieve efficiency. In most cases, Large Deviations Theory provides the criteria for the appropriate modification of the simulation density. The presentation comprises the motivation for variance reduction and some introductory aspects on Large Deviations and Importance Sampling. Then the analysis of a class of open-loop digital phase modulation receivers is presented. Additive white Gaussian noise and random carrier phase are assumed in the model.

THE VITERBI AND BCJR ALGORITHMS WITH SIGNAL DEPENDENT CORRELATED NOISE José M. F. Moura Professor, Dept. of Electrical and Computer Engineering, Carnegie Mellon University, USA (on sabbatical leave at MIT) 15:30 h, Thursday, July 27, 2000 Room: Anfiteatro EA2 ABSTRACT
We consider the design of sequence detectors for channels with intersymbol interference (ISI) and correlated (and/or signal-dependent) noise. We describe three major contributions: (i) First, by modeling the noise as a finite order Markov process, we derive the optimal Maximum-Likelihood sequence detector (MLSD) and the optimal maximum a posteriori MAP sequence detector, extending to the correlated noise case the Viterbi and BCJR algorithms. We show that, when the signal dependent noise is conditionally Gauss-Markov, the branch metrics in the MLSD are computed from the conditional second order noise statistics. We evaluate the branch metrics using a bank of finite impulse response (FIR) filters. (ii) Second, we characterize the error performance of the MLSD and MAP sequence detectors. The error analysis of these detectors is complicated by the correlation asymmetry of the channel noise. We derive upper and lower bounds and computationally efficient approximations to these bounds based on the banded structure of the inverses of Gauss-Markov covariance matrices. An experimental study shows the tightness of these bounds. (iii) Finally, we derive several classes of suboptimal sequence detectors, and demonstrate how these and others available in the literature relate to the MLSD. We compare their error rate performance and their relative computational complexity, and show how the structure of the MLSD and the performance analysis guide us in choosing a best compromise between several types of suboptimal sequence detectors. In the final part of the talk we will comment on extensions of the work to soft decoding, and turbo/LDPC codes with correlated, possibly signal-dependent noise.

BAYESIAN LEARNING OF MODEL STRUCTURE Zoubin Ghahramani Professor, Gatsby Computational Neuroscience Unit, University College London, England 14:30 h, Tuesday, June 27, 2000 ISR's Meeting Room ABSTRACT
In any field where models are built from data there is a tension between fitting the data well and keeping the model simple. I will discuss this problem in the context of learning the structure of latent variable models and other probabilistic graphical models encountered in pattern recognition and machine learning. The Bayesian approach allows a principled treatment of the problem of learning model structure. The tension between data fit and model complexity is resolved via Ockham's razor, which arises from averaging over all possible settings of the model parameters. Unfortunately, for most non-trivial problems these averages are intractable, resulting in the use of large-sample limits, local Gaussian approximations, or Markov chain Monte Carlo methods. I will describe an approach to Bayesian inference based on variational approximations. Variational methods are deterministic, global, generally fast, and the objective function is guaranteed to increase monotonically. The variational optimisation procedure generalises the EM algorithm. The optimal forms of the approximating distributions fall out of the optimisation (and need not be Gaussian). Most importantly, the variational Bayesian approach can be used to compare and adapt model structures since the objective function transparently incorporates the model complexity cost. This approach can be used (1) to automatically infer the most probable number of clusters and the intrinsic latent-space dimensionality of each cluster, (2) to find the number of sources in a blind source separation problem, (3) to infer the dimensionality of the state space of a linear dynamical system, and (4) to optimise over the structure of a mixtures of experts network.

SORTING CONTINUOUS-TIME SIGNALS: THE ANALOG MEDIAN FILTER Paulo J.S. Ferreira Professor, Universidade de Aveiro / IEETA 14:30 h, Wednesday, May 24, 2000 ISR's Meeting Room ABSTRACT
The median filter, ever since Tukey's suggestion in the early seventies, has been at the center of vigorous research in nonlinear filtering. Many variants and generalizations of the basic running median have been proposed, leading to a rich class of nonlinear operators with interesting properties (median filters preserve steps and sharp transitions while removing impulsive noise). Virtually all the literature on order-statistics and related filters concerns discrete-time filters and signals, a fact that is hardly surprising: there is no conceptual difficulty in sorting the samples of a discrete-time signal, but it is less obvious how to devise a similar operation for continuous-time signals.

This introductory talk explores the fact that continuous-time (even multidimensinal) signals can be meaningfully ``sorted'', or rearranged, leading to precise formulations of analog median filters and other variants of order-statistics filters. The properties of the rearrangements are briefly discussed, bearing in mind their role in the analysis of a class of nonlinear filters, of which the median filter is the most well-known example. It is hoped that the new perspective might contribute to a deeper understanding of the properties and implementation of these nonlinear filters.

ROBUST AND ADAPTIVE FILTERING BASED ON THE QQ-PLOT APPROACH Zeljko Djurovic Post-Doc, Instituto de Sistemas e Robótica Instituto Superior Técnico and Professor of Electrical Engineering University of Belgrade February 24, 2000 ABSTRACT
There are many engineering applications, such as radar, sonar or communication systems, where the presence of measurement noise can decrease significantly the efficiency of the applied signal processing algorithms. In some of them it is necessary to design estimation algorithm insensitive to the bad measurements, i.e., outliers. Such algorithms are said to be robust. On the other hand, many algorithms have to be adaptive due to the change of measurement noise statistical properties. In this talk, two different approaches to the design of robust or adaptive filters and estimation procedures will be presented.

The QQ-plot is a statistical technique that relies on a specific graphical representation of the samples of a stochastic process, which is generally used to verify the prior assumption about the process distribution function. However, the QQ-plot contains much more information about the statistical properties of the stochastic process. Firstly, the statistics of the first and second order can be estimated based on it. Also based on the shape of the QQ-plot, it is even possible to design appropriate algorithms for probability density function estimation. Finally, the idea of robust filtering may found an adequate geometrical explanation, and robust influence function result as a natural solution from the QQ-plot framework.

It is well known that accurate estimation of the probability density function of a stochastic process, based on statistical approaches, usually requires a huge number of its samples. An alternative approach relies on pattern recognition methods to recognize the actual probability density function from the apriori-adopted classes. Depending on the size of the measurement noise record, the probability of correct decision may be very high. This information can be used in order to tune the processing functions in adaptive filtering algorithms. In this way, the estimation error variance can be decreased significantly.

UNCERTAINTY IN THE TIME - FREQUENCY PLANE Paulo Mónica de Oliveira Ph.D. Student, Instituto de Sistemas e Robtica Instituto Superior Técnico and Navy Shcool September 30, 1999 ABSTRACT
Is there a limit to the maximum resolution one can achieve when representing the signal's energy in the Time-Frequency plane? Some authors sustain that such a limit exists, and ignoring it is the cause of the known difficulties with some joint Time-Frequency distributions. Others maintain that there is no such limit. We propose to analyze the merits and demerits of the several approaches, and suggest further arguments one might wish to consider. This will take us to the conclusion that, both from a tool-specific and from a general information theoretic point of views, there is indeed a lower limit on the achievable resolution, even though the expression of that limit can not be given by the traditional Heisenberg-Gabor relations.

A CLASS OF DISTRIBUTED DECISION PROBLEMS Michael Athans Visiting Professor, Instituto de Sistemas e Robotica Instituto Superior Tecnico and Professor of Electrical Engineering (Emeritus) Massachussetts Institute of Technology October 9, 1998 ABSTRACT
We discuss a class of truly distributed optimal decision problems that arise in hypothesis-testing and binary detection problems. We investigate optimal team-theoretic decision rules under the assumption that several decision agents within an organization make independent uncertain measurements of the same event, coordinate using constrained communication protocols, and arrive at a team-optimal decision. These problems are much more complex than their centralized counterparts and demonstrate subtle issues of decision making in decentralized settings. We examine the improvement in the quality of the decision as we allow increased communication among the decision agents. In addition, we discuss issues of different decision-oriented organizational architectures, e.g. flat vs. hierarchical, and the interplay between organizational topologies and restricted communication protocols for coordination. The results illustrate why it is difficult to structure superior (expanding or contracting) organizations.