Evolution of neural networks, or neuroevolution, has been a successful approach to many low-level control problems such as pole balancing, vehicle control, and collision warning. However, certain types of problems — such as those involving strategic decision-making — have remained difficult to solve. This dissertation proposes the hypothesis that such problems are difficult because they are fractured: The correct action varies discontinuously as the agent moves from state to state. To evaluate this hypothesis, a method for measuring fracture using the concept of function variation of optimal policies is proposed. This metric is used to evaluate a popular neuroevolution algorithm, NEAT, empirically on a set of fractured problems. The results show that (1) NEAT does not usually perform well on such problems, and (2) the reason is that NEAT does not usually generate local decision regions, which would be useful in constructing a fractured decision boundary. To address this issue, two neuroevolution algorithms that model local decision regions are proposed: RBF-NEAT, which biases structural search by adding basis-function nodes, and Cascade-NEAT, which constrains structural search by constructing cascaded topologies. These algorithms are compared to NEAT on a set of fractured problems, demonstrating that this approach can improve performance significantly. A meta-level algorithm, SNAP-NEAT, is then developed to combine the strengths of NEAT, RBF-NEAT, and Cascade-NEAT. An evaluation in a set of benchmark problems shows that it is possible to achieve good performance even when it is not known a priori whether a problem is fractured or not. A final empirical comparison of these methods demonstrates that they can scale up to real-world tasks like keepaway and half-field soccer. These results shed new light on why constructive neuroevolution algorithms have difficulty in certain domains and illustrate how bias and constraint can be used to improve performance. Thus, this dissertation shows how neuroevolution can be scaled up from learning low-level control to learning strategic decision-making problems. / text

Streptococcus pyogenes (Group A Streptococcus) is a Gram-positive bacterial pathogen that causes significant superficial and invasive diseases. Iron acquisition is an important component of GAS pathogenesis in the human host. The 10 gene sia operon of GAS is involved in the acquisition of iron via heme or heme-binding proteins and encodes an ABC transporter as well as the large multifunctional receptor Shr. Domain analysis of Shr shows that it contains two copies of the DUF1533 (domain of unknown function) in its N-terminal part and two NEAT (NEAr Transporter) domains. NEAT domains are found in variable copy number in surface proteins of Gram-positive pathogens and are implicated in binding to various ligands. A new recombinant Shr protein was cloned and a purification protocol was developed, improving the yield of the full-length protein. A solid phase binding assay was developed and used to demonstrate Shr binding to hemoglobin. Several truncated Shr proteins were expressed and purified: the N-terminal Domain (NTD) up to but not including the first NEAT domain of Shr, the NTD plus the first NEAT domain (NTD-NEAT1) and the second NEAT domain alone (NEAT2). It was determined that Shr’s NTD mediates hemoglobin binding, demonstrating that a new protein pattern in Shr is involved in hemoglobin binding, and implicating the DUF1533 in this process. It was also determined that NTD-N1 and NEAT2 bind heme while NTD does not. Therefore, both NEAT domains may participate in the capture of heme from the host hemoglobin by Shr.

Los Robots Modulares Reconfigurables son sistemas que han despertado un gran interés científico durante los últimos años. Ellos corresponden a un caso especial de Materia Programable, donde es posible cambiar la forma y función del sistema modular mediante, por ejemplo, la reconfiguración de sus partes. Los sistemas robóticos modulares están constituidos por diversas unidades capaces de obtener información autónomamente, procesándola de forma distribuida o centralizada. Algunas de las interrogantes fundamentales de esta área de investigación son cómo distribuir la información y toma de decisiones entre los distintos módulos o cómo contribuir, mediante procesamiento y actuación local, al comportamiento global del sistema robótico; entre otras. Si bien otros investigadores han desarrollado estrategias de control para generar comportamientos simples (locomoción unidireccional, formación de geometrías simples, auto reparación), no existen en la literatura métodos que permitan la generación automática de dichas reglas. Para esto, se desarrolló un simulador acoplado a distintos métodos de aprendizaje evolutivo para la síntesis de algoritmos de locomoción para robots modulares reconfigurables genéricos. Los métodos de aprendizaje empleados fueron Algoritmos Genéticos y NEAT (Neuro Evolution of Augmenting Topologies). De esta forma, con algoritmos genéticos se desarrollaron reglas de locomoción explícitas (similares a las disponibles en la literatura), mientras que con NEAT se desarrolló una novedosa propuesta, donde las reglas de locomoción están contenidas implícitamente en redes neuronales. Finalmente, se estudió el proceso de obtención de los algoritmos de locomoción y, posteriormente, se caracterizaron las distintas soluciones obtenidas ante la variación de distintos parámetros y escenarios de ensayo. El error asociado al aprendizaje con NEAT fue de 1.93%, mientras que con Algoritmos Genéticos se tuvo un error de 17.87%, de donde se tiene que el entrenamiento con NEAT es más repetible que con Algoritmos Genéticos. Sin embargo, los resultados finales obtenidos con Algoritmos Genéticos superaron a los obtenidos con NEAT en un 17.88%. Para mejorar el proceso de entrenamiento fue necesario acondicionar el simulador de forma tal que cada algoritmo de locomoción obtuviese siempre el mismo resultado, reduciendo la aleatoridad.

Thesis describe theory behind neuroevolution. Then it describes both design and creation of simulated environment for autonomous agent and its training with library Neataptic in environments with various difficulty. Thesis also describes pro- cess of designing frontend for visualization of results and backend for faster training of agents. At the end it describes resulting agents and proposes enhancements to existing solution.

Dans l'état actuel des techniques de détection des exoplanètes, aucune planète tellurique du Système Solaire ne pourrait être détectée et pourtant leur présence est une contrainte très forte sur les scénarios de formation des systèmes planétaires. L'astrométrie, en mesurant l'effet reflex des planètes sur leur étoile centrale, permet de remonter à la masse des planètes et aux paramètres orbitaux. C'est une technique très utilisée pour la détermination des masses et des orbites des étoiles binaires et couronnée de succès. Il est nécessaire d'aller dans l'espace pour atteindre les précisions nécessaires pour détecter toutes les planètes jusqu'aux masses telluriques. Le Laboratoire est engagé dans un projet qui a été proposé à l'ESA dans le cadre de l'appel à mission M3 de Cosmic Vision et qui a pour objectif de recenser toutes les planètes de notre voisinage solaire. Le principe est d'utiliser l'astrométrie différentielle pour compléter les mesures obtenues par les autres techniques afin de descendre le seuil de détection et de caractérisation au niveau de la masse terrestre dans la zone habitable de chaque système. Nous voulons explorer de façon exhaustive toutes les étoiles de type solaire (type spectral FGK) jusqu'à 20pc de notre Soleil. Le concept du satellite repose en l'état actuel de l'étude sur du vol en formation avec un satellite portant le miroir et un satellite portant le plan focal. La mesure est faite par une métrologie à base interférométrique.Le sujet de la thèse consiste à avancer d'une part sur la définition du cas scientifique et d'autre part sur la spécification de l'instrument et des procédures d'observation. En ce qui concerne le cas scientifique, il s'agit d'établir une stratégie d'observation optimale pour recenser et caractériser de manière exhaustive tous les systèmes planétaires de notre voisinage. A l'aide de simulations numériques, l'étudiant pourra établir une stratégie de réduction des données permettant de remonter aux paramètres des orbites planétaires. Il s'agira aussi de participer à l'établissement du budget d'erreur de l'instrument et à la définition des modes d'observation. L'étudiant sera aussi mené à mettre en œuvre des tests de performance de la mesure dans le cadre d'une étude expérimentale. Cette thèse se déroulera dans le cadre de la collaboration européenne sur le sujet et des contacts seront aussi tissés avec nos collègues du JPL qui maitrisent la métrologie. / With the present state of exoplanet detection techniques, none of the rocky planets of the Solar System would be detected and indeed their presence is a very strong constraint on the scenarios of the formation of planetary systems. Astrometry by measuring the reflex effect of planets on their central host stars, lead us to the mass of planets and to their orbit determination. This technique is used frequently and is very successful to determine the masses and the orbits of binary stars. However it is necessary to go to space to reach the precision required to detect all planets down to the telluric regime.We are proposing a mission to ESA in the framework of the call for M3 mission in the Cosmic Vision plan whose objective is to make a full census of all exoplanets in our Solar neighborhood. The objectif is to use differential astrometry to complete the measurements obtained by other techniques in order to lower the threshold of detection and characterization to the level of an Earth mass in the habitable zone of each system. We want to explore in an exhaustive manner all solar-type stars (FGK spectral type) up to 20pc from the Sun. The satellite concept is based on formation flying technology with a satellite carrying a single primary mirror and another satellite carrying the focal plane. The measure is done using laser metrology using interferometry.The topic of the thesis consists in making progress on the definition of the science case and on the specification of the instrument and the observing procedures. Concerning the science case, an optimized observing strategy has to be defined to exhaustively detect and characterize all planetary systems in the solar neighborhood. Using numerical simulations, the student will establish a strategy for data reduction that allows him to fit all orbital parameters of the systems. A participation to the computation of error budget of the instrument and to the definition of observing modes is expected too. The candidate can also carry out performance tests using existing testbeds or developing new ones. This thesis will take place in the framework of the European collaboration on this topic and contacts will be made with our JPL colleagues who master the metrology technique.

The bacterium Streptococcus pyogenes requires heme, which is taken up via an ABC transporter. An understanding of this pathway may result in new approaches to antibacterial agents. Both SiaA and NEAT2 (NEAr Transporter 2) are proteins involved in heme binding. One of the axial ligands of SiaA, His 229, was purified to study how mutagenesis affects heme binding. UV-visible studies showed a small band at 420 nm with respect to the protein band at 288 nm which probably indicates that heme was lost easily from this mutant. We have also worked to optimize the yield of Shr-NEAT2 by changing different variables. For each of the batches, the yield of holoNEAT2 was calculated by UV-visible spectroscopy. Increasing oxygen during growth did not improve holoNEAT2 yield. On the other hand, lower temperature, decrease in time after induction, and addition of ALA all increased the protein production.

Group A streptococcus (GAS) is an important pathogen that produces a wide spectrum of suppurative infections and autoimmune sequelae in humans, ranging from less complex pharyngitis, impedigo to more severe manifestations such as necrotizing fasciitis, toxic shock syndrome, rheumatic fever and glomerulonephritis. The worldwide burden of GAS infections and sequelae is considerable, but an immunization program that defends against the hyper-variable GAS is missing. The streptococcal hemoprotein receptor (Shr), is an iron-regulated protein involved in heme acquisition. An unspecified region in the amino terminus of Shr mediates the interactions with hemoglobin and two protein modules named NEAT1 and NEAT2 bind heme. In this study, we analyzed the molecular structure and function of Shr, investigated its antigenic properties and role in GAS disease production. We demonstrated that Shr is a new type of GAS adhesin that contributes to the pathogen interactions with extracellular matrix (ECM) proteins. Shr enabled bacterial adherence to host cells and was important for GAS virulence in vivo. Immunizations with Shr protein by intraperitoneal or intranasal administration conferred resistance to systemic GAS challenge in mice. Shr antiserum allowed bacterial opsonization and defended against GAS diseases in a murine model for passive vaccination. Studies with isolated Shr domains localized ECM-binding to the NEAT domains and showed that most of the protein is exposed on the bacterial surface. In addition, Shr N-terminal region and both of the NEAT modules elicited strong antibody response in rabbits. In conclusion, Shr is a protective antigen that contributes to GAS pathogenesis by facilitating both heme uptake and bacterial adherence. Since Shr is conserved among GAS strains and other pyogenic streptococci, this study demonstrates that Shr may be used to develop a vaccine against GAS strains and related pathogens.

AN ABSTRACT OF THE THESIS OF SOROOSH SOHANGIR, for the MASTER OF SCIENCE degree in COMPUTER SCIENCE, presented on 9 th November 2011, at Southern Illinois University Carbondale. TITLE: OPTIMIZED FEATURE SELECTION USING NEUROEVOLUTION OF AUGMENTING TOPOLOGIES (NEAT) MAJOR PROFESSOR: Dr. Shahram Rahimi Feature selection using the NeuroEvolution of Augmenting Topologies (NEAT) is a new approach. In this thesis an investigation had been carried out for implementation based on optimization of the network topology and protecting innovation through the speciation which is similar to what happens in nature. The NEAT is implemented through the JNEAT package and Utans method for feature selection is deployed. The performance of this novel method is compared with feature selection using Multilayer Perceptron (MLP) where Belue, Tekto, and Utans feature selection methods is adopted. According to unveiled data from this thesis the number of species, the training, accuracy and number of hidden neurons are notably improved as compared with conventional networks. For instance the time is reduced by factor of three.

Doctor of Philosophy / Biochemistry and Molecular Biophysics Interdepartmental Program / Phillip E. Klebba / Gram-positive bacteria are characterized by a single lipid bilayer with a thick peptidoglycan layer. This group of organisms contains bacteria commonly associated with human infection, including: Staphylococcus aureus, Listeria monocytogenes, Bacillus anthracis and Streptococcus pneumoniae among others. These bacteria have a common system for importing iron in the form of heme, which is acquired by proteins containing heme-binding NEAT (NEAr iron Transporter) domains. The heme acquisition system in S. aureus is termed the Iron Surface Determinant (Isd) system and in L. monocytogenes is termed Heme Binding Protein (Hbp) and Heme/Hemoglobin Uptake Protein (Hup). These proteins work together to obtain heme from hemoglobin and then transport the heme into the cytoplasm via well characterized ABC-transporters. Although there have been clinical trials with antibodies directed against Isd proteins, there are currently no antibiotics targeting iron uptake systems in bacteria in general. Building upon fluorescent approaches for detection of iron uptake in Gram negative organisms, this work develops fluorescent heme acquisition detection in Gram positive organisms. The spectrofluorimetric methodology facilitates the understanding of heme acquisition protein interactions and mechanisms in bacteria. This work could subsequently be used to identify inhibitors of Gram positive bacterial iron uptake systems, and develop a new target for antibiotic action.

Context. Interest in and research on neural networks and their capacity for finding solutions to nonlinear problems has increased greatly in recent years. Objectives. This thesis attempts to compare competitive coevolution to traditional neuroevolution in the game StarCraft: Brood War. Methods. Implementing and evolving AI-controlled players for the game StarCraft and evaluating their performance. Results. Fitness values and win rates against the default StarCraft AI and between the networks were gathered. Conclusions. The neural networks failed to improve under the given circumstances. The best networks performed on par with the default StarCraft AI.