IDSIA Seminars

Davide Nitti - Boolean Description Logic Games for automated multi-attribute negotiation

Tue, 7 Dec 2010 12:00:35 +0100

Automated negotiation has recently received significant attention due to the development of the World Wide Web and especially of the Semantic Web. Indeed, these technologies has provided the necessity for the further development of computational negotiation and bargaining techniques. In this talk, I propose an implementation of Boolean Description Logic Games: an automated multi-attribute negotiation combining Boolean Games with ontological background knowledge, formulated in expressive description logics. I begin by describing how Boolean Games work, after I discuss how to extend Boolean Games with the power of description logics, and finally, I show a real implementation (made for my Master’s thesis) of Boolean D.L. Games using java, with examples and performance evaluation.
This is a join work with Thomas Lukasiewicz, Azzurra Ragone and Tommaso Di Noia

Kamil Ciosek - New Developments in Semi-supervised Learning

Fri, 12 Nov 2010 12:00:49 +0100

Recently, a great deal of work within machine learning has gone into the semi-supervised paradigm, which consists in using unlabeled data samples in the hope that additional information on the distribution of training objects improves classification or regression performance. This talk gives a quick overview of semi-supervised versions of two popular kernel algorithms: support vector machines and Gaussian processes as well as discusses a result for arbitrary RHKSs. For support-vector machines, an interpretation wrt. the VC bound is given, whereas for Gaussian processes a Bayesian justification for semi-superivsed learning is provided. An effort is made to discuss the relative merits and weaknesses of each of the approaches, particularly with respect to computational complexity.

Prof. Laura Pozzi - Protein Structure Prediction in the HP (Hydrophobic-Polar) Lattice Model

Mon, 4 Oct 2010 11:30:52 +0200

In this talk, I will present an existing simple exact model, called the HP (Hydrophobic-Polar) model, for protein structure prediction. I will then describe the algorithm I have worked on, based on pruned exponential search, to solve the problem exactly in the 2D lattice.
The algorithm currently finds optimal solutions for protein sequences of around 50-60 elements

A. Michael Maye - Invariants and symmetries of perception: How problems of information processing may relate to Autism

Mon, 6 Sep 2010 12:00:40 +0200

The basic idea of this work is to investigate analogies between clinical phenomena as Asperger Syndrome and Autism Spectrum Disorders (ASD) on one hand and problems in state of the art autonomous agents in dynamic environments on the other hand. Individuals with ASD show a variety of atypical behavioral and perceptual patterns which seem to lack a common underlying cause. We hypothesize that there may be a general impairment of the autistic brain characterized by a missing restriction on the class of connectivity patterns, i.e. features, that can be utilized for learning perceptual and cognitive tasks. We
suggest that the autistic brain may not make proper use of symmetries and develop invariants to permutations which significantly enhances the computational requirements. The idea is illustrated by a simple model. Examples of behavioral characteristics of ASD children are discussed and it is investigated in which way they can be explained by our model.

I. Fasel and M. Asada have contributed to this work

Dr. Thomas Rothvoss - An improved LP-based approximation for the Steiner tree problem

Tue, 20 Jul 2010 11:15:11 +0200

The Steiner tree problem is one of the most fundamental NP-hard problems:
given a weighted undirected graph G=(V,E) and a subset of terminal nodes, find a minimum weight tree spanning the terminals.
In this talk, we introduce a directed LP relaxation and describe an approximation algorithm based on iterative randomized rounding of a fractional solution.
We prove an approximation guarantee of 1.39 for the algorithm (improving over the previously best known factor of 1.55) and show that the mentioned LP has an integrality gap of at most 1.55 (improving over the previously best known factor of 2).

Karen Yeressian - A simple model of bounded rationality

Thu, 17 Jun 2010 12:00:29 +0200

In the talk a rather simple mathematical model of bounded ra- tionality which is inspired by the simulated annealing algorithm is presented. Having such a model we may measure the rationality of an agent with respect to some utility function. First we consider the problem of decision under un- certainty then an abstract definition of the cost of rationality and at the end we briefly consider the multiple target (utility) case and multiple agent games.

Dr Easter Selvan - ICA Learning Exploiting Riemannian Geometry

Thu, 17 Jun 2010 12:00:20 +0200

A set of stochastic algorithms to optimize the independence criterion for implicit imposition of the orthonormality constraint among the estimated sources will be presented. The major advantage of the proposed algorithms is the increased accuracy with which the weight matrix in the independent component analysis (ICA) model is estimated, compared to conventional schemes. In a pursuit to relax the orthonormality constraint, a design of the steepest descent (SD), conjugate gradient (CG) and quasi-Newton (QN) methods, intended for oblique manifold, will also be discussed with a few experimental results.

Athanasios Kalpakas - Modeling and Verifying an Innovative Biomass-based Thermal Energy System

Tue, 15 Jun 2010 11:30:09 +0200

Biomass exploitation systems are able to supply heat, steam and electricity for home or industrial users. However, modern utilization methods of biomass do not deal with the problem globally but usually get confined to empirical and non-optimal approaches. The innovative energy production unit which is designed and implemented in this paper suggests a universal processing and exploitation strategy using several different types of biomass while aiming at an effective combustion with minimum loses. System identification and modeling of the combustion process is implemented using a simplified linear model which can adequately describe the above process and lay strong foundations for a prospective robust control.

Saša Nešić - Semantic Document Architecture for Desktop Data Integration and Management

Mon, 14 Jun 2010 11:00:29 +0200

In this talk I will present a novel desktop document architecture, namely SDArch, which attempts to integrate data from heterogeneous application-specific document formats into a unified desktop information space. SDArch introduces a novel document representation model, called the Semantic Document Model (SDM), which integrates the semantic layer into the core of the document data representation structures. The model is inline with the Linked Data principles, representing documents as RDF linked data sets, in which document units are identified by globally unique URIs, semanticaly annotated by an arbitrary number of ontological concepts and interlinked via explicit semantic links. Semantic documents have only one permanent instance, which is integrated into a unified desktop data graph and stored into a desktop RDF repository. The RDF repository also exposes a publicly available SPARQL endpoint, thus enabling remote access to the desktop data and its integration into Linked Open Data (LOD) cloud the global Semantic Web.

On top of the desktop RDF repository, SDArch introduces a service-oriented middleware with a set of services such as the semantic search and navigation services, which aim to improve the search and navigation within desktop data, thus improving the effectiveness and efficiency of desktop users in carrying out their daily tasks. Together, the RDF repository, the middleware and the application software that utilizes the middleware services compose SDArch as a three-tier service-oriented architecture.

We conducted a usability evaluation of the SDArch prototype and an experimental evaluation of the semantic document search. The evaluation results were promising. I will also briefly present the analysis of these results.

Dr Davide Eynard - SAnno framework for Semantic Annotation

Tue, 1 Jun 2010 11:30:58 +0200

When browsing the Web, you would often like to know what people think about the site you are visiting; sanno in Italian means "they know", and the main goal of this project is to let people share semantic annotations about any resource reachable on the Web. Filtering and browsing annotations coming from distributed and heterogeneous sources is made possible by a general and flexible model based on named graphs; while the vocabulary about annotation metadata is well defined and fixed, any RDF property can be used to describe a Web resource, though the convergence on shared vocabularies is boosted.
The framework is designed to be distributed, based on the fact that information is provided in a specific semantic format and users can subscribe only to the annotation sources they like and/or trust.
The system relies on user participation but can be bootstrapped thanks to the conversion of already existing structured or semi-structured annotation systems. Semantic annotations can finally be shown through a browser extension, which relies on predefined visualization techniques to manage different families of properties.

Ms Agnieszka Jastrezsbka - Virtualization technologies on the large and small scale: IBM mainframe solutions and VMware Server and Workstation products.

Fri, 28 May 2010 11:30:40 +0200

This talk presents some theoretical and empirical aspects of virtualization technologies. First I describe the basic techniques of hardware partitioning, bare-metal hypervisor and hosted hypervisor. Then I will report on the virtualization techniques designed for implementation on System z mainframe platform, focussing on LPARs and z/VM operating system, their architecture and principles of operation, highlighting how mainframes can guarantee almost continuous availability and deliver high performance. In particular we will examine the z/VM system environment with SUSE Linux Enterprise Server guest machine that IBM Poland made kindly available for this Master Thesis. Then I will introduce VMware Server and Workstation products shortly discussing the differences between the server and the workstation environments and providing a short demo. The results are obtained using monitoring the Linux Fedora core using the nmon analyzer tool. Finally I will discuss how virtualisation impacts business investment choices.

Dr Nic Schraudolph - New Inference Methods in Binary Markov Random Fields

Thu, 27 May 2010 12:00:25 +0200

Reweighted Weighted Matching is my new algorithm for computing a ground state of an Ising spin glass, resp. (equivalently) a maximum graph cut, resp. a MAP state of a pariwise binary Markov random field. It uses a Langrangian relaxation to leverage the duality between Ising states and perfect matchings from planar to nonplanar graphs, where the ground state problem is known to be NP-hard and non-approximable. Nonetheless my algorithm always runs in polynomial time, and is much faster and scales to much larger graphs than the best previously known methods. The question of P=NP is preserved by the possibility of our algorithm returning a "don't know" answer, though this has never yet been observed in best-practice use. This strongly suggests that the max cut problem is polynomial in the average and perhaps even smoothed worst case, on graphs where it is worst-case NP-hard.

Ilaria Vacca - The Tactical Berth Allocation Problem: integrated optimization in container terminals

Wed, 19 May 2010 11:00:11 +0200

In the context of container terminal operations, the simultaneous optimization of decision problems that are usually solved hierarchically by terminal's planners represents nowadays a promising research trend. In this talk we introduce the Tactical Berth Allocation Problem (TBAP), that deals with the integration of the berth allocation problem (BAP) and the quay crane assignment problem (QCAP). The objective is to schedule incoming ships over a time horizon, assigning them a berthing position and a certain quay crane profile (i.e. number of quay cranes per working shift). The problem has been modeled as a mixed integer program and housekeeping costs generated by the berth assignment are taken into account by a quadratic term in the objective function. The added value of the integrated optimization approach will be analyzed and illustrated
by numerical examples based on real-world data.

Linus Gisslen - Why are perylene pigments colored?

Tue, 11 May 2010 12:00:26 +0200

Perylene derivatives are robust organic dyes absorbing and emitting light in the visible range and in the near infrared. Perylene and its derivatives have attracted significant interest as active layers for light harvesting, photovoltaics, and photoinduced charge and energy transfer processes.
In order to increase their efficiency and stability, these materials have been studied quite intensively. In particular, their possible application in solar-cells needs insight into optical excitation and charge transport processes. The performance of these materials as charge or energy transport materials does not arise exclusively from the electronic properties of the individual molecules, but it depends as well on favorable intermolecular interactions, such as π-stacking.
This talk will give a brief introduction to these materials and an insight in the processes behind their optical properties. These properties are fundamental to understanding why perylene materials have the colors that they have.
by numerical examples based on real-world data.

Dr Ioannis Dokas - Socio-technical Early Warning Systems: A case study from the Republic of Ireland

Thu, 6 May 2010 11:00:52 +0200

Socio-technical systems are those which rely not only on technology but on humans and social organizations for their adequate functioning. Failures in these systems, depending on the magnitude of their adverse affects, can lead to disasters. It is widely accepted that proactive risk management strategies is a way towards reducing the numbers of disasters in socio-technical systems. In this context early warning systems and their services can be seen as tools that support these strategies.
In my talk I will introduce the "Socio-technical Early Warning System" concept and I will present some results of our research that aims at developing a web based socio-technical early warning system for the water utilities in the Republic of Ireland. This novel early warning system is intended to be used by many actors within different organisations dedicated to the provision of clean, wholesome drinking water in the Republic of Ireland, such as the Irish Environmental Protection Agency and the Health Service Executive.

Professor Ferdinando Villa - Semantic meta-modelling: towards true modularity and interoperability in model specification

Tue, 27 Apr 2010 11:00:46 +0200

Building and running models of natural and artificial systems is a complex activity that has so far defied most attempts to facilitate the interoperability of paradigms, contexts, and technologies beyond a basic level of generality. True modularity in model specifications, so that models can be independently developed, communicated and used across contexts of application, has remained a holy grail in modelling practice because of the difficulty of maintaining the semantic and mathematical integrity of models assembled from independent components.
I will describe a novel formalism for modeling, based on the explicit statement of the semantics of observables and observations, that allows consistent specification and true interoperability for models of many formerly incompatible types. Probabilistic vs. deterministic, spatially explicit vs. aggregated, agent-based vs. process-based are some of the
distinctions that semantic meta-modelling allows bridging. After discussing the fundamentals of the approach and briefly illustrating the modeling formalism and its implementation in open source software, I will demonstrate
an application to a web-based environmental decision support system.

Jürgen Leitner - Controlling Spacecraft with Artificial Neural Networks

Mon, 26 Apr 2010 12:00:27 +0200

This talk will focus on the Advanced Concepts Team (ACT) of the European Space Agency and its current research projects, mainly focusing on the areas of Artificial Intelligence and Robotics.
The ACT has been working on evolving Artificial Neural Networks (ANNs) for space applications as a first step to add autonomy to future spacecraft operations. These neurocontrollers can be seen as black- boxes translating inputs (position, attitude and velocities) to control outputs (thrusters), substituting hand- coded control laws. In this paper we show that evolved ANNs can be used to control one or more spacecrafts in an environment simulating zero friction and tidal gravity.
The first application considered is the autonomous docking of a spacecraft. Autonomous docking is defined as one spacecraft approaching another, with hard constraints on the final position, the final speed (usually very close to zero) and the final attitude of the spacecraft. An ANN representation is used and evolved using PaGMO, an open source software platform for global optimization of unconstrained problems developed by the ACT.
This talk tries to explain the problems of autonomous docking, the approach chosen by the ACT and will present current research and results.

Fabrizio Patuzzo - A bio-inspired robot controller

Mon, 26 Apr 2010 11:30:14 +0200

Developing controllers for biped robots is one of the great challenges of 21st century robotics. This research is expected to find applications ranging from service robotics and entertainment robotics to the exploration of terrains inaccessible to humans (atomic nuclear plants, distant planets, etc.). But it is also useful to other areas like prosthetics (to develop prostheses for handicapped people), sports science (to improve athlete’s performance), or neuroscience (to better understand the motor control system in animals).
This short talk presents a controller for biped robots inspired by how vertebrate animals produce locomotion. Three ingredients allow vertebrate animals to walk: clusters of neurons in the spinal cord; control commands from the brain stem; and sensory feedback. One can model this scheme with coupled non linear oscillators, and use this mathematical model to design a controller for a humanoid robot.

Dr Paola Villa - KALC: a constructive semantics for ALC

Tue, 20 Apr 2010 11:30:03 +0200

In this talk we firstly present a Kripke semantics for the description logic ALC which is directly inspired by the semantics for Intuitionistic logic. Moreover, we discuss why a direct translation of this kind of semantics is not adequate in the description logic context and propose a constructive semantics that differs from the previous one by the fact that we impose a condition on the partial order. We also present a tableau calculus which is sound and complete with respect to our semantics. As an application of the calculus we prove that this semantics meets the disjunction property, a key criterion used in assessing whether a logic is constructive.

Dr Andrea Fossati - 3–D Reconstruction from a Single View Input

Tue, 20 Apr 2010 11:00:15 +0200

Obtaining a 3–D output from a single-view, 2–D input is an intrinsically ill-constrained problem. Therefore additional appropriate information is needed to overcome this issue and produce a mean- ingful and realistic 3–D reconstruction. I will show how this concept is valid in interesting and concrete computer vision problems, like human body tracking, vehicle tracking and 3–D reconstruc- tion of buildings.