Augsburg, Univ., Diss., 2008.

In this chapter we studied a first approach to generate suitable rule sets for solving classification problems on systems of autonomous, memory constrained components. It was shown that a multi agent system that uses interacting Pittsburgh-style classifier systems can evolve appropiate rule sets. The system evolves specialists for parts of the classification problem and cooperation between them. In this way the components overcome their restricted memory size and are able to solve the entire problem. It was shown that the communication topology between the components strongly influences the average number of components that a request has to pass until it is classified. It was also shown that the introduction of communication costs into the fitness function leads to a more even distribution of knowledge between the components and reduces the communication overhead without influencing the classification performance very much. If the system is used to generate rule sets to solve classification tasks on real hardware systems, communication cost in the training phase can thus lead to a better knowledge distribution and small communication cost. That is, in this way the system will be more robust against the loss of single components and longer reliable in case of limited energy resources.

A Sustainable Autonomic Architecture for Organically Reconfigurable Computing System based on SRAM Field Programmable Gate Arrays (FPGAs) is proposed, modeled analytically, simulated, prototyped, and measured. Low-level organic elements are analyzed and designed to achieve novel self-monitoring, self-diagnosis, and self-repair organic properties. The prototype of a 2-D spatial gradient Sobel video edge-detection organic system use-case developed on a XC4VSX35 Xilinx Virtex-4 Video Starter Kit is presented. Experimental results demonstrate the applicability of the proposed architecture and provide the infrastructure to quantify the performance and overcome fault-handling limitations. Dynamic online autonomous functionality restoration after a malfunction or functionality shift due to changing requirements is achieved at a fine granularity by exploiting dynamic Partial Reconfiguration (PR) techniques. A Genetic Algorithm (GA)-based hardware/software platform for intrinsic evolvable hardware is designed and evaluated for digital circuit repair using a variety of well-accepted benchmarks. Dynamic bitstream compilation for enhanced mutation and crossover operators is achieved by directly manipulating the bitstream using a layered toolset. Experimental results on the edge-detector organic system prototype have shown complete organic online refurbishment after a hard fault. In contrast to previous toolsets requiring many milliseconds or seconds, an average of 0.47 microseconds is required to perform the genetic mutation, 4.2 microseconds to perform the single point conventional crossover, 3.1 microseconds to perform Partial Match Crossover (PMX) as well as Order Crossover (OX), 2.8 microseconds to perform Cycle Crossover (CX), and 1.1 milliseconds for one input pattern intrinsic evaluation. These represent a performance advantage of three orders of magnitude over the JBITS software framework and more than seven orders of magnitude over the Xilinx design flow. Combinatorial Group Testing (CGT) technique was combined with the conventional GA in what is called CGT-pruned GA to reduce repair time and increase system availability. Results have shown up to 37.6% convergence advantage using the pruned technique. Lastly, a quantitative stochastic sustainability model for reparable systems is formulated to evaluate the Sustainability of FPGA-based reparable systems. This model computes at design-time the resources required for refurbishment to meet mission availability and lifetime requirements in a given fault-susceptible missions. By applying this model to MCNC benchmark circuits and the Sobel Edge-Detector in a realistic space mission use-case on Xilinx Virtex-4 FPGA, we demonstrate a comprehensive model encompassing the inter-relationships between system sustainability and fault rates, utilized, and redundant hardware resources, repair policy parameters and decaying reparability.

As reconfigurable devices' capacities and the complexity of applications that use them increase, the need for self-reliance of deployed systems becomes increasingly prominent. A Sustainable Modular Adaptive Redundancy Technique (SMART) composed of a dual-layered organic system is proposed, analyzed, implemented, and experimentally evaluated. SMART relies upon a variety of self-regulating properties to control availability, energy consumption, and area used, in dynamically-changing environments that require high degree of adaptation. The hardware layer is implemented on a Xilinx Virtex-4 Field Programmable Gate Array (FPGA) to provide self-repair using a novel approach called a Reconfigurable Adaptive Redundancy System (RARS). The software layer supervises the organic activities within the FPGA and extends the self-healing capabilities through application-independent, intrinsic, evolutionary repair techniques to leverage the benefits of dynamic Partial Reconfiguration (PR). A SMART prototype is evaluated using a Sobel edge detection application. This prototype is shown to provide sustainability for stressful occurrences of transient and permanent fault injection procedures while still reducing energy consumption and area requirements. An Organic Genetic Algorithm (OGA) technique is shown capable of consistently repairing hard faults while maintaining correct edge detector outputs, by exploiting spatial redundancy in the reconfigurable hardware. A Monte Carlo driven Continuous Markov Time Chains (CTMC) simulation is conducted to compare SMART's availability to industry-standard Triple Modular Technique (TMR) techniques. Based on nine use cases, parameterized with realistic fault and repair rates acquired from publically available sources, the results indicate that availability is significantly enhanced by the adoption of fast repair techniques targeting aging-related hard-faults. Under harsh environments, SMART is shown to improve system availability from 36.02% with lengthy repair techniques to 98.84% with fast ones. This value increases to "five nines" (99.9998%) under relatively more favorable conditions. Lastly, SMART is compared to twenty eight standard TMR benchmarks that are generated by the widely-accepted BL-TMR tools. Results show that in seven out of nine use cases, SMART is the recommended technique, with power savings ranging from 22% to 29%, and area savings ranging from 17% to 24%, while still maintaining the same level of availability.

L’Intelligence Ambiamte (AmI) est un domaine de recherche investigant les techniques d’intelligence artificielle pour créer des environnements réactifs. Les réseaux de capteurs et effecteurs sans-fils sont les supports de communication entre les appareils ménagers, les services installés et les interfaces homme-machine. Cette thèse s’intéresse à la conception d’Environements Réactifs avec des propriétés autonomiques i.e. des systèmes qui ont la capacité de se gérer eux-même. De tels environements sont ouverts, à grande échelle, dynamique et hétérogène, ce qui induit certains problèmes pour leur gestion par des systèmes monolithiques. L’approche proposée est bio-inspirée en considérant chacune des plate-formes comme une cellule indépendente formant un organisme intelligent distribué. Chaque cellule est programmée par un processus ADN-RNA décrit par des règles réactives décrivant leur comportement interne et externe. Ces règles sont modelées par des agents mobiles ayant des capacités d’auto-réécriture et offrant ainsi des possibilités de reprogrammation dynamique. Le framework EMMA est composé d’un middleware modulaire avec une architecture orientée ressource basée sur la technologie 6LoWPAN et d’une architecture MAPE-K pour concevoir des AmI à plusieurs échelles. Les différentes relations entre les problèmes techniques et les besoins théoriques sont discutées dans cette thèse depuis les plate-formes, le réseau, le middleware, les agents mobiles, le déploiement des applications jusqu’au système intelligent. Deux algorithmes pour AmI sont proposés : un modèle de contrôleur neuronal artificiel pour le contrôle automatique des appareils ménagers avec des processus d’apprentissage ainsi qu’une procédure de vote distribuée pour synchroniser les décisions de plusieurs composants systèmes. / AThe Ambient Intelligence (AmI) is a research area investigating AI techniques to create Responsive Environments (RE). Wireless Sensor and Actor Network (WSAN) are the supports for communications between the appliances, the deployed services and Human Computer Interface (HCI). This thesis focuses on the design of RE with autonomic properties i.e. system that have the ability to manage themselves. Such environments are open, large scale, dynamic and heterogeneous which induce some difficulties in their management by monolithic system. The bio-inspired proposal considers all devices like independent cells forming an intelligent distributed organism. Each cell is programmed by a DNA-RNA process composed of reactive rules describing its internal and external behaviour. These rules are modelled by reactive agents with self-rewriting features offering dynamic reprogramming abilities. The EMMA framework is composed of a modular Resource Oriented Architecture (ROA) Middleware based on IPv6 LoW Power Wireless Area Networks (6LoWPAN) technology and a MAPE-K architecture to design multi-scale AmI. The different relations between technical issues and theoretical requirements are discussed through the platforms, the network, the middleware, the mobile agents, the application deployment to the intelligent system. Two algorithms for AmI are proposed: an Artificial Neural Controller (ANC) model for automatic control of appliances with learning processes and a distributed Voting Procedures (VP) to synchronize the decisions of several system components over the WSAN.