Acevedo-Valle, J. M.; Trejo, K.; Angulo, C. International Conference of the Catalan Association for Artificial Intelligence p. 196-205 DOI: 10.3233/978-1-61499-806-8-196 Data de presentació: 2017-10-25 Presentació treball a congrés
Within the machine learning framework, incremental learning of multivariate spaces is of special interest for on-line applications. In this work, the regression problem for multivariate systems is solved by implementing an efficient probabilistic incremental algorithm. It allows learning high-dimensional redundant non-linear maps by the cumulative acquisition of data from input-output systems. The proposed model is aimed at solving prediction and inference problems. The implementation introduced in this work allows learning from data batches without the need of keeping them in memory afterwards. The learning architecture is built using Incremental Gaussian Mixture Models. The Expectation-Maximization algorithm and general geometric properties of Gaussian distributions are used to train the models. Our current implementation can produce accurate results fitting models in real multivariate systems. Results are shown from testing the algorithm for both situations, one where the incremental learning is demonstrated and the second where the performance to solve the regression problem is evaluated on a toy example.
La publicació definitiva d'aquest treball està disponible a IOS Press a través de http://dx.doi.org/10.3233/978-1-61499-806-8-196
Angulo, C.; Acevedo-Valle, J. M. International Conference of the Catalan Association for Artificial Intelligence p. 46-55 DOI: 10.3233/978-1-61499-806-8-46 Data de presentació: 2017-10-25 Presentació treball a congrés
According to the sensorimotor approach, cognition is constituted by regularities in the perceptual experiences of an active and situated agent. This perspective rejects traditional inner representational models, stressing instead patterns of sensorimotor dependencies. Those relations are called sensorimotor contingencies (SMC). Many research areas and accounts are working on and related with it. In particular, four distinct kinds of SMCs have been previously introduced for environment, habitat, coordination and strategy using dynamical models from a psychological perspective. As dynamical systems, in this paper we analyze SMCs, for the very first time, from a modern control engineering perspective. We provide equations and block diagrams translating the psychological proposal to control engineering. We also analyze the original toy example proposed from the psychological domain into the modern control engineering point of view, as well as we propose a first approach to this toy example coming from the control engineering domain.
This work introduces new results on the modeling of early-vocal development using artificial intelligent cognitive architectures and a simulated vocal tract. The problem is addressed using intrinsically-motivated learning algorithms for autonomous sensorimotor exploration, a kind of algorithm belonging to the active learning architectures family. The artificial agent is able to autonomously select goals to explore its own sensorimotor system in regions where its competence to execute intended goals is improved. We propose to include a somatosensory system to provide a proprioceptive feedback signal to reinforce learning through the autonomous discovery of motor constraints. Constraints are represented by a somatosensory model which is unknown beforehand to the learner. Both the sensorimotor and somatosensory system are modeled using Gaussian mixture models. We argue that using an architecture which includes a somatosensory model would reduce redundancy in the sensorimotor model and drive the learning process more efficiently than algorithms taking into account only auditory feedback. The role of this proposed system is to predict whether an undesired collision within the vocal tract under a certain motor configuration is likely to occur. Thus, compromised motor configurations are rejected, guaranteeing that the agent is less prone to violate its own constraints.
The present work focuses on two main objectives. Firstly, it highlights the relevance of studying the early stages of language development using machines as an approach to contribute to the future of speech recognizers and synthesizers, user interfaces, active learning techniques, and to the field of robotics and artificial intelligence in general. Secondly, this work introduces some results on the study of the role of somatosensory models in vocal autonomous exploration. In previous works, the roles of intrinsic motivations and motor constraints in early vocal development were studied showing that active learning techniques can be used by artificial agents endowed with a simulated vocal tract to autonomously learn how to produce intended sounds through the use of probabilistic models. This work studies the effects of modifying the somatosensory model, which is used to map motor commands to undesired articulatory configurations, over the intrinsically motivated active learning process. The somatosensory system is modeled as a Gaussian Mixture Model. Herein, some simulations were run varying the structure of the model in order to analyze differences in the results. The effects on the explored sensorimotor regions and the amount of undesired vocal configurations are studied. The simulations presented in this work show that the structure of the current somatosensory model is relevant to the learning process. However, it can be also concluded that in order to reliably characterize the effects of modifying the somatosensory model further simulations must be performed and clear measures for performance should be considered. // El trabajo presentado persigue dos objetivos principales: el primero de ellos es mostrar la necesidad de estudiar las etapas tempranas del desarrollo del lenguaje utilizando máquinas. Estos estudios contribuirán en el desarrollo futuro de sintetizadores y reconocedores de voz, interfaces de usuario e indirectamente al estudio de la inteligencia artificial; el segundo objetivo es presentar nuevos resultados en el estudio sobre el rol de los sistemas somatosensores en la exploración vocal temprana. En trabajos preliminares fueron estudiados los roles de las motivaciones intrínsecas y las restricciones motoras en el desarrollo vocal temprano. De estos estudios se concluyó que las técnicas de aprendizaje automático activo pueden ser utilizadas en conjunto con agentes artificiales dotados con un tracto vocal simulado para aprender autónomamente cómo producir sonidos específicos. En el presente trabajo se estudian los efectos del cambio de los parámetros que definen el modelo probabilístico del sistema somatosensorial, el cual mapea configuraciones motoras con configuraciones articulares indeseadas sobre el proceso de aprendizaje. El sistema somatosensorial es modelado utilizando “Gaussian Mixture Models”. A través del resultado de una serie de simulaciones donde se modifica la estructura del modelo antes mencionado, se demuestra que la estructura del modelo somatosensorial es relevante para el proceso de aprendizaje. Sin embargo, los resultados también indican que para realizar una mejor caracterización de los efectos de la modificación del modelo somatosensorial deben llevarse a cabo más simulaciones, así como tomar en consideración nuevas medidas de calidad del aprendizaje.
Acevedo-Valle, J. M.; Puig, V.; Tornil-Sin, S.; Witczak, M.; Rotondo, D. IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes p. 1-6 Data de presentació: 2015-09-02 Presentació treball a congrés
The present work proposes a Fault Tolerant Control (FTC) methodology for non-linear discrete-time systems that can be modeled as Linear Parameter Varying (LPV) systems. The proposed approach relies on the modeling of faults as additional scheduling parameters of the LPV model for the controlled system and it uses a triple loop architecture. The inner control loop is designed by means of the standard H2/H1 control methodologies based on Linear Matrix Inequalities (LMIs). The design takes into account a prespecified set of faults and the ranges of their magnitudes that are wanted to be tolerated and it assumes the availability of on-line fault estimations provided by a Fault Detection and Isolation (FDI) module. The resulting controller tries to compensate the system faults in order to maintain a satisfactory closed-loop dynamic performance, but it does not take into account possible system input and state constraints associated to actuator saturation and other physical limitations. Thus, an intermediate control loop determines the actual compensation feasibility using set invariance theory. And, when it is needed, it applies suitable additive predictive control actions that enlarge the invariant set, trying to assure that the current state remains inside the enlarged invariant set. Finally, an outer loop implements a model reference control that allows reference tracking. The use of the proposed FTC methodology is illustrated through its application to the well-known quadruple tank system benchmark.
The four-tanks system is well-known and a considerable number of works study it in the existent literature. This work proposes a nonlinear model predictive control (NMPC) design to regulate the water ¿ow inputs of the four tanks system. As an initial approach, the implementation is performed in Simulink prior to a foreseeable real-plant application of the proposed solution. Simulation results are presented and discussed to show the performance of the controller. Improvements are enumerated and analysed.
Rotondo, D.; Puig, V.; Acevedo-Valle, J. M.; Nejjari, F. International Conference on Control and Fault-Tolerant Systems p. 492-497 DOI: 10.1109/SysTol.2013.6693844 Data de presentació: 2013-10-10 Presentació treball a congrés
In this paper, an FTC strategy using Linear Parameter Varying (LPV) virtual sensors is proposed and applied to the IFAC wind turbine case study. The novelty of the proposed strategy consists in that virtual sensors are applied to the FTC problem in a new original fashion. Instead of hiding the fault, the virtual sensors are used to expand the set of available sensors. Then, the state observer is designed using LPV techniques based on Linear Matrix Inequalities (LMIs) taking into account a varying parameter that is introduced in order to select which sensors are used by the observer among the physical and the virtual ones. In this sense, the proposed approach can be considered as a multisensor fusion strategy that integrates data provided by various sensors in order to obtain a better estimation.