A foundation for perception in autonomous systems = Fundamentos de la percepción en sistemas autónomos

Abstract

Esta tesis doctoral ofrece una visión conceptual de la percepción, con la intención de ser aplicada en el futuro al análisis y al diseño de sistemas artificiales. Tal y como está expuesta en este trabajo, ofrece un marco conceptual que permite extraer principios y reglas generales sobre el funcionamiento de los sistemas. El marco construido no está formalizado, por lo que a ún no permite una cuantificación. La formalización y la cuantificación son dos pasos que deberían seguir a este trabajo >cap.3. Se observará que la tesis tiene dos capítulos. El primero, cap. 1, está dedicado a los sistemas autónomos. Tiene por objeto explicar conceptos abstractos tales como la autonomía, y cómo se relacionan con la operación y la estructura de los sistemas. Se comienza aportando una visión de los sistemas artificiales existentes desarrollada en torno al concepto de autonomía. A continuación, se exponen los conceptos más importantes de la teoría general de sistemas tal y como se entenderá en este trabajo. A partir de aquí, se desarrolla una visión de los sistemas autónomos y generales que trata de integrar aspectos internos (como su estructura) con otros externos (como su autonomía o su comportamiento). Este contexto determina el papel de la percepción en los sistemas autónomos y la forma en que tiene lugar. Ilustra las relaciones de la percepción con los demás aspectos sistémicos y las restricciones potenciales a las que está sometida. El segundo capítulo (cap. 2) está dedicado a la percepción propiamente dicha. Como se decía, se entiende que la mayoría de sus aspectos estructurales se derivan directamente del contexto sistémico desarrollado en el capítulo anterior. En línea con esta idea, este capítulo analiza la percepción desde dos puntos de vista que derivan del concepto de sistema: Por una parte, las partes que intervienen en la percepción y la manera en que están relacionadas. Por otra parte, los flujos de información en el sistema. Finalmente, en el capítulo 3 se incluye una exposición con cierto nivel de profundidad explicando las principales conclusiones y líneas de progreso previstas para este trabajo. La versión española de esta tesis es un resumen de la versión inglesa. Es esta ´ ultima la que debe consultarse. La versión española está construida sintetizando la mayoría de conceptos y las explicaciones de la inglesa. Secciones enteras han sido excluidas del re- inglesa se ha respetado con el fin de facilitar al lector referirse a los contenidos de esta versión. Los objetivos principales de esta tesis son los siguientes: Generalidad: Explicar la percepción desde un punto de vista general, estableciendo una ontología común para sistemas artificiales y biológicos. Obtener conceptos, principios y relaciones aplicables al diseño de sistemas artificiales. Este objetivo incluía una formalización de la ontología. Estos objetivos se formularon desde el convencimiento de que los niveles de complejidad y la naturaleza de las tareas de los sistemas artificiales actuales exceden los niveles de prestaciones ofrecidos por la ingeniería convencional. La generalidad, eventualmente, permitirá diseñar soluciones bioinspiradas eficientes a problemas todavía sin resolver. El diseño bioinspirado ha existido en la ingeniería desde tiempos remotos. Ejemplos de aproximaciones recientes de este tipo pueden encontrarse en arquitecturas cognitivas conocidas como RCS y SOAR. Este trabajo se enmarca en una línea de investigación centrada principalmente en la investigación sobre teorías y principios generales más que sobre problemáticas y aplicaciones concretas, aunque éstas formen parte de ella necesariamente. Su foco es la ingeniería del conocimiento, aplicada a cualquier faceta del diseño de sistemas. La generalidad es una condición necesaria para ésto. El interés por la formulación general de problemas no es nuevo. Su expresión más clara se dio con el nacimiento de la teoría general de sistemas, general system theory, GST a mediados del siglo XX. De hecho, este trabajo se basa en conceptos sobre sistemas heredados de una de las formulaciones de esta teoría, recogida en el libro An Approach to General Systems Theory, por George J. Klir [Kli69].3 El grado en que se ha alcanzado los objetivos expuestos arriba se discutirá en un capítulo al final de esta versión espa˜ nola, y más en detalle en su homólogo de la versión inglesa. Sin embargo, es conveniente avanzar que no se ha completado al formalización y que solo se ha alcanzado un grado elemental. La metodología diseñada para llevar a cabo este trabajo se ha basado en el ideal del método científico tradicional, que podemos resumir en un ciclo de tres fases fundamentales: (1) experimentación (2) observación y (3) generalización. De acuerdo con este ideal, la experimentación sirve tanto como punto de partida como referente frente al que comprobar las nuevas teorías. Debido al alto grado de multidisciplinaridad de este trabajo y el interés por la generalidad, fue necesario reinterpretar el ideal para hacer la investigación posible con tiempo y recursos limitados. La fase de ‘experimentación’ se reformuló en un análisis experimental, en en cual se analizó tanto la experiencia previa del Grupo en sistemas inteligentes de control como multitud de fuentes externas al grupo, y de diversas áreas. Llevar a cabo experimentos de psicología, neurociencia, ingeniería, geometría y todas las demás disciplinas en que se ha basado este trabajo hubiera sido imposible. Se ha realizado un gran esfuerzo para extraer principios generales de los trabajos anteriores del Grupo y de literatura científica, que pudiesen aportar casuística de referencia y cumplir el mismo papel que la experimentación propiamente dicha. Esta interpretación del método científico, adoptada al principio de la investigación, la hemos entendido como esencial más tarde, a medida que sumábamos más áreas de conocimiento a aquellas de las que partimos: percepción en sistemas biológicos. A la hora de poner por escrito el trabajo realizado en los ´ utlimos a˜ nos, ha sido necesario restringir el punto de vista del discurso para hacerlo más comprensible y darle coherencia. Por ello, se ha omitido diversos temas que fueron importantes para llegar a la conceptualización propuesta aquí. Entre ellos se cuentan la consciencia, los sistemas paralelos distribuidos, los sistemas de tiempo real, estudios sobre arquitecturas cognitivas, y otras disciplinas, en menor grado, como la geometría. Es conveniente se ˜ nalar, por ´ ultimo, que la mayoría de los conceptos expuestos aquí, de acuerdo con los objetivos expuestos arriba, son generales y tienen un alto nivel de abstracción. Por tanto, deben explicar tanto lo complejo como lo simple, lo concreto y lo abstracto, lo artificial y lo biológico de una manera coherente. En los sistemas reales, muchos de los conceptos que aquí se mencionan pueden no aparecer, o hacerlo de una forma muy primitiva, mientras que otros pueden darse de forma muy desarrollada. No existe sistema alguno conocido por el autor que desarrolle plenamente todos los conceptos expuestos aquí. In accordance with the title of this work, the main objective is to build a conceptual foundation for perception in autonomous systems. This derived in two major goals: Generality: Explaining perception from a general point of view, establishing a common ontology for artificial and biological systems. Obtaining concepts, principles and relations applicable to artificial system design and to artificial perceptive system design. This objective included a formalization of the ontology. These objectives were stated under the belief that the levels of complexity and the nature of the tasks needed by current artificial systems exceeded the level of performance enabled by conventional engineering. Generality would eventually permit applying efficient bioinspired solutions to currently unsolved technical problems. Bioinspiration has existed in engineering since ancient times; perhaps the most known example is the study of the wings of birds in order to build flying artefacts. More recently and related to this work, specific approaches have led to biologically inspired cognitive architectures, of which RCS [Alb99], [Alb95], [GMP+01] and SOAR [New90], [RLN93], [LBCC99] are perhaps the best known and inspiring. This work falls within a line of research which aims at general principles and theories more than at specific projects and problems. Its focus is engineering knowledge applicable to any problem regarding autonomous system design. Generality is a necessary condition for this. The aim for generality is not novel. In the recent history of science, the interest for generality, relations, sets and isomorphisms experienced a progressive rise during the XIX century and perhaps reached is zenith in the mid XX century. The title General System Theory [vB69] is regarded as the foundation of the theory centered, precisely, in the study of systems as to their systemhood, regardless their circumstantial features: i.e. General Systems. This way of understanding science and reality, promising as it remained for decades, seemed abandoned at the beginning of this work. In fact, it was not until well advanced in this research that the Theory of General Systems was adopted as the background for the investigation, in the particular formulation of An Approach to General Systems Theory [Kli69]. The degree up to which the objectives have been achieved is discussed in chapter 15. However, it is worth advancing that a complete formalization has not been proposed, and that this work provides only a semi-formal discourse. The methodology designed for this work followed the ideal of the traditional scientific method, conceptualized in figure 15.1, p.246, consisting of three major phases: (1) experimentation, (2) observation and (3) generalization, in which experimentation serves both as the starting point and as the benchmark against which to check generalizations. Due to the multidisciplinary nature of this work and the aim for generality, it was necessary to restate the ideal in order to make the investigation possible in finite time and with finite resources. The stage of ‘experimentation’ was reformulated into a thorough analysis of the experience of the Group1 in intelligent control system design, and an extensive bibliographic research in this and other fields, trying to cover all the scope from experimental research to abstract theories. Performing experiments on psychology, neuroscience, engineering, geometry and all other disciplines related to this work would have been impossible. Instead, it was decided to carry out a major effort in order to transform the documental corpus into both a source of general principles and concepts, and an experimental benchmark against which to test them. This interpretation of the scientific method, envisaged at the beginning of this research, proved essential later, when more and more new fields of knowledge were added to that which started the research: perception in biological systems. Some knowledge domains studied have been mentioned explicitly in the text, while others contributed to form the concepts proposed, but have been left implicit: consciousness, geometry, art, algebra among others. Among the second, we must remark that consciousness, in all the perspectives considered during this research [Anc99], [Baa97], [Den91], [Hol03], [Lyo95], [Tay99], parallel distributed systems and real–time systems [BW97], [Jal94], [MR86], [RM86], [Sch95] and other miscellaneous sources [Fra95], [KD95], [New90], [Ame99] contributed to developing the distributed conception of systems which has been adopted in this work. The thesis has been structured following a general to particular scheme. The topics which are treated first are the contextual, which are follwed by the more specific ones. Firstly, a discourse on systems establishes the context for perception. Then perception is developed. This is followed by an analysis of real systems. The work concludes with a discussion about the major achievements and conclusions and reference material. The same general to particular scheme has been applied within each part, assessing contextual aspects in the first chapters and progressively entering the more specific ones. Part II is dedicated specifically to systems, in order to offer a general vision of them, which are the context in which perception takes place. Concepts of distributed systems, general systems and engineering are integrated in a unified notion. This discourse on systems has intended to prepare the reader for a clear, straightforward discourse on perception which, in other case, would have proved excessively interleaved with systemic considerations. Chapter 5 offers a short study of autonomy, the problems involved in building autonomous systems, and the different ways in which it has been approached. In this light, it offers an overview of artificial systems addressing their different strategies for autonomous behaviour. Finally it explores some fundamental aspects related with autonomy in systems. Chapter 6 offers an overview of the major theoretical and methodological source of this work: An Approach to General Systems Theory, by George J. Klir, [Kli69]. It introduces the main concepts and ideas which will be used throughout the text. Chapter 7 is the main exposition of this part. It integrates multiple concepts about systems inherited from many sources in a unified vision. The internal aspects about systems such as their structure are related with external ones such as behaviour and autonomy. This chapter develops the systemic framework in which perception will be explained in part III. Part III is dedicated to perception. It develops the topic in the context given by the system, in the terms introduced in part II, understanding that the reader should conceive the concepts and perceptive processes within the restrictions and dynamics of a systemic context: environment, objectives, resources, distributed functions, perturbances, etc. Chapter 8 describes the problem of perception from a global, basic perspective in order to identify the major parts and processes to be explained. It then describes a collection of relevant approaches to perception from this perspective, indicating the exact aspects of the problem in which each study is focused. Chapter 9 states the main points of this thesis schematically. Chapter 10 takes the discourse on perception from the introduction to the fundamentals of problem of chapter 8 and develops it into a detailed view of the process and its parts. Chapter 11 develops the informational or cognitive aspects of perception, establishing a relation with the taxonomical analysis of chapter 10. Part IV analyzes examples of real systems in detail, in order to illustrate the concepts of parts II and III. Chapter 12 describes an embedded automotive system for detecting losses of attention in the driver. It is analyzed in detail mainly for aspects on perception. Chapter 13 describes the CONEX system, an example of a complex intelligent control system, a past development of the ASLab Group. It is analyzed both for systemic and perception concepts. This chapter was contributed by the director of this work, Dr. Ing. Ricardo Sanz, who actually took part in the CONEX project. Chapter 14 analyzes the case of a fault-tolerant, massively parallel system for concepts like reconfiguration, adaptivity and functional decomposition, which are not easily found in artificial systems. Part V includes chapter 15, in which the major achievements of this work are discussed, and the resulting framework compared to existing theories. Part VI Includes the initial versions of a glossary on general autonomous systems and specialized terms of perception, and the list of bibliographic references. Finally, it is worth remarking, although it will be outlined throughout the text, that the concepts introduced here are general. This means that they must explain the simple and the complex, the particular and the abstract, the natural and the artificial. No real system is known to the author to fully develop the generality of all the concepts proposed in this work.2 In real systems, some of the aspects mentioned here may appear in such a primitive form as to be only ‘degenerated’ instances of our concepts. Others may appear to fully develop the notions proposed here. One of the objectives of the line of research in which this work has emerged is to be able to design systems in which their characteristics are developed up to an arbitrary degree, at the choice of the designer. Of all characteristics, autonomy would be perhaps the most tempting, and initiates the discourse. Mastering the design of parts and structure of systems will eventually enable this. However, we shall deduce from the text that this objective is so ambitious that it may well become a dream. Let this work be a primitive step in the way to this dream.