Giuliano Armano

Intelligent Information Access

The book aims to investigate current developments and new insights into methods, techniques and technologies for intelligent information access from a multidisciplinary perspective. It comprises six chapters authored by participants in the research event Intelligent Information Access, held in Cagliari (Italy) in December 2008.

NXCS Experts for Financial Time Series Forecasting

Most of the early machine learning approaches -e.g., Decision Lists (DL) [1], [2], Decision Trees (DT) [3], Counterfactuals (CFs) [4], Classification And Regression Trees (CART) [5]- apply the divide-and-conquer principle by recursively partitioning the input space until regions of roughly constant class membership are obtained. The corresponding algorithms yield a monolithic result by enforcing heuristics devised to control the complexity of the search. Notwithstanding this apparent interpretation, they can also be reviewed in the light of a new perspective, in which the partitioning procedure is considered as a tool for generating multiple experts. Although with a different focus, both the evolutionary-computation and the connectionist communities allowed to make explicit the multiple experts’ perspective. In the former community, the focus was on establishing suitable architectures and techniques able to enforce an adaptive behavior on a population of individuals, e.g., Genetic Algorithms (GAs) [6], [7], Learning Classifier Systems (LCSs) [8], [9], and eXtended Classifier Systems (XCSs) [10]. In the latter community, the focus was mainly on training techniques and outputs combination mechanisms; in particular, let us recall Jordan and Jacobs’ Mixtures of Experts (MEs) [11], [12] and Weigend’s Gated Experts (GEs) [13].

A rapid development process with UML

Since design-oriented life-cycles came to their maturity, dramatic changes have been introduced as far as programming tools and computer hardware are concerned. Such changes made it possible to develop applications focusing on refactoring rather than on analysis and design. The underlying hypothesis is that by adopting suitable tools and target languages, refactoring would possibly cost less than the overhead introduced by modern A&D techniques. Recently, extreme programming has been proposed as an alternative to a design-oriented life-cycle.In this paper we describe a software application developed using a software life-cycle that basically follows the guidelines suggested by extreme programming. Such an approach requires highly expressive programming languages and powerful CASE tools. UML has been selected as the underlying modeling language throughout the whole process, for it incorporates well-known diagrams for describing a software application from different perspectives. Smalltalk has been selected as target language, as it allows fast prototyping and early delivery. We claim that, for small and medium-sized projects, a life-cycle based on refactoring and supported by suitable languages and tools allows team productivity to be greatly enhanced.

Mixture of random prototype-based local experts

The Mixture of Experts (ME) is one of the most popular ensemble methods used in pattern recognition and machine learning This algorithm stochastically partitions the input space of the problem into a number of subspaces, experts becoming specialized on each subspace The ME uses an expert called gating network to manage this process, which is trained together with the experts In this paper, we propose a modified version of the ME algorithm which first partitions the original problem into centralized regions and then uses a simple distance-based gating function to specialize the expert networks Each expert contributes to classify an input sample according to the distance between the input and a prototype embedded by the expert As a result, an accurate classifier with shorter training time and smaller number of parameters is achieved Experimental results on a binary toy problem and selected datasets from the UCI machine learning repository show the robustness of the proposed method compared to the standard ME model.

Why software agents can be effective in biomedical sciences

Nowadays most of the software systems, tools, and applications in biomedical sciences are very complex and require high flexibility. Seemingly, it is difficult to identify a common underlying technology. In fact, each specific research field imposes its own constraints on the corresponding solutions. Nevertheless, at least from a software engineering perspective, the agent technology appears to fit the requirements imposed by such different research fields and encompasses most of the biomedical systems, tools, and applications in a unifying framework.

Computational Design of a DNA- and Fc-Binding Fusion Protein

Computational design of novel proteins with well-defined functions is an ongoing topic in computational biology. In this work, we generated and optimized a new synthetic fusion protein using an evolutionary approach. The optimization was guided by directed evolution based on hydrophobicity scores, molecular weight, and secondary structure predictions. Several methods were used to refine the models built from the resulting sequences. We have successfully combined two unrelated naturally occurring binding sites, the immunoglobin Fc-binding site of the Z domain and the DNA-binding motif of MyoD bHLH, into a novel stable protein.

Planning by Abstraction Using HW

This paper presents the parametric system HW[], devised and implemented to perform planning by abstraction. The parameter is an external PDDL-compliant planner, which is exploited to search for solutions at any required level of abstraction including the ground one. To represent abstraction hierarchies an extension to the standard PDDL notation has been devised, which allows one to control communication between adjacent levels by specifying suitable translation rules. Moreover, a novel semi-automatic technique for generating abstract spaces is also described. Five domains -taken from the AIPS 2002, 2000 and 1998 planning competitions- have been selected as test bed. To assess the impact of abstraction techniques on search performances, experiments have been focused on comparing three relevant planners with their hierarchical counterparts. Results put into evidence that the hierarchical approach may significantly speed up the search.

Special issue: Multi-agent systems for medicine, computational biology, and bioinformatics

Recent progresses in the area of biomedical sciences, including medicine, bioinformatics and computational biology has made it essential to devise new insights and technological platforms able to ensure appropriate support to research activities. To facilitate the most relevant tasks, a new generation of systems, tools, and applications needs to be developed, aimed at ensuring the feasibility of significant advances in all relevant subfields of biomedicine. Medical and biological information is increasing at an impressive rate, and it is still typically made available to researchers through web interfaces – which are actually mainly suitable to human beings rather than to software applications. Doctors and biologists that work on the border line of state-of-the-art research topics are accustomed to connect to many sites characterized by heterogeneous interfaces and data structures, as well as to switch among them in order to submit and to retrieve preliminary and intermediate data. It is clear that no generic description can be given, able to encompass all requirements imposed by different disciplines and kinds of researchers, nevertheless –in all cases– researchers must deal with complex systems and envision a kind of “entropic change”, with the feeling that a new era for systems, tools, and applications is approaching. Note that “complex”, here, does not necessarily address computational problems. Rather, the complexity lies in the conceptualization, analysis, and design of the system, tool, or application to be implemented, as well as in the interactions among entities (despite their “nature” of objects, components, or agents) that are involved at run-time. Several cross-sectional issues occur in the cited biomedical sciences. For the sake of brevity, let us consider only: (i) the problem of integrating heterogeneous information and resources, (ii) the problem of simulating and modelling complex biological systems, and (iii) the problem of providing relevant specific tools and applications, possibly via web interfaces. As for the integration of heterogeneous information and resources, any solution should be given keeping in mind that an essential requirement is the openness of the proposal – to ensure that newly structured information can be processed, provided that a suitable ontological description of it is given. As for the simulation and modeling of complex biological systems (see [Bosse et al.], [Bandini et al.], [Merelli and Young]), which are crucial issues in systems biology and computational biology, it is clear the need for specific tools and notations. The need for distribution (in order to speed-up the corresponding computation), and for the adoption of software paradigms able to deal with concurrencyrelated issues are further relevant concerns. As for the need of specific tools and applications, usually supplied through web interfaces and finalized to problemsolving, visualization, or other analytical tasks (see [Bortolussi and Dovier]), it is clear that they must undergo a rapid change in order to support new technologies, e.g. web services, and to provide results according to a holistic view (see [Koutkias et al.]). Although the path forward is still long and difficult, the agent technology is now being assessed in biomedicine, also for its ability of dealing with information integration, distribution, and with concurrencyrelated issues. It is worth pointing out that the paradigmatic change issued by the adoption of this new technology would not imply a dramatic change, as agents can encompass existing applications, through suitable wrappers, as well as new technologies, such as web services, through suitable gateways.

Mutual implications and granularity

Abstract This paper illustrates a technique for discovering mutual implications among hierarchically structured data. Such a technique may be applied to both knowledge and data bases. If the hierarchical structure makes it possible to define granularity levels, mutual implications can be evaluated at any level. Results can be quantitative (i.e. a degree in the range [0, 1]) or qualitative (i.e. a label taken from a user-defined set). If the ground data do not represent a mapping among individuals, i.e. the level of information granularity is not the highest, a local approximation based on T -Norms can be used. The process of implication discovery allows one to derive inference rules for expert systems and to detect default values. In addition, it might be successfully used by sophisticated machine learning algorithms.