Intelligence computationnelle | EPFL Graph Search

Lintelligence computationnelle (IC) est un domaine scientifique. Il n'existe pas de définition admise par tous les auteurs, mais c'est un domaine lié à l'intelligence artificielle. Parmi les méthodes utilisées dans ce domaine, la plupart des auteurs comptent les heuristiques, la logique floue, les réseaux de neurones et les algorithmes évolutionnistes. Une définition controversée Il n'existe pas de définition majoritaire pour « intelligence computationnelle ». En particulier son lien avec l'intelligence artificielle n'est pas le même pour tous les auteurs. Pour certains c'est le même domaine, pour d'autres l'intelligence computationnelle est une branche de l'intelligence artificielle et enfin certains définissent l'intelligence computationnelle comme une nouvelle approche de l'intelligence artificielle. Axes et outils La plupart des auteurs, ainsi que la Société d'Intelligence computationnelle (« Computational Intelligence Society » (CIS)) de l'IEEE établie en

Comparison Among Computational Intelligence Methods For Engine Knock Detection. Part 2

Anamaria Radoi

The novelty this article brings is the use of neural networks and neuro-fuzzy modeling to take the engine knock detection based on pressure and vibration samples taken from an internal combustion engine to the next step from previous achievements, the ultimate goal being higher detection rates and also faster response times than the classical non neuro-fuzzy methods. Work started from the theoretical works available in the domain, algorithms were adapted for the case in hand, then the study was led on how they would be affected by the variety of situations that occur in internal-combustion engines with the scope of real-time applications. Following the experiments, results were finally compared showing significantly greater knock detection rate-time improvements than other methods employed so far.

2012

Reduced Basis method for Isogeometric Analysis: application to structural problems

Marina Rinaldi

Isogeometric Analysis (IGA) is a computational methodology for the numerical approximation of Partial Differential Equations (PDEs). IGA is based on the isogeometric concept, for which the same basis functions, usually Non-Uniform Rational B-Splines (NURBS), are used both to represent the geometry and to approximate the unknown solutions of PDEs. Compared to the standard Finite Element method, NURBS-based IGA offers several advantages: ideally a direct interface with CAD tools, exact geometrical representation, simple refinement procedures, and smooth basis functions allowing to easily solve higher order problems, including structural shell problems. In these contexts, repeatedly solving a problem for a large set of geometric parameters might lead to high and eventually prohibitive computational cost. To cope with this problem, we consider in this work the Reduced Basis (RB) method for the solution of parameter dependent PDEs, specifically for which the NURBS representation of the computational domain is parameter dependent. RB refers to a technique that enables a rapid and reliable approximation of parametrized PDEs by constructing low dimensional approximation spaces. In this work, for the construction of the reduced spaces we adopt two different strategies, namely the Proper Orthogonal Decomposition and the greedy algorithm. In this thesis we combine RB and IGA for the efficient solution of parametrized problems for all the possible cases of NURBS geometrical parametrizations, which specifically include the NURBS control points, the weights, and both the control points and weights. In particular, we first focus on the solution of second order PDEs on parametrized lower dimensional manifolds, specifically surfaces in the three dimensional space. We consider geometrical parametrizations that entail a nonaffine dependence of the variational forms on the spatial coordinates and the geometric parameters. Thus, depending on the parametrization at hand and in order to ensure a suitable Offline/Online decomposition between the reduced order model construction and solution, we resort to the Empirical Interpolation Method (EIM) or the Matrix Discrete Empirical Interpolation Method (MDEIM), by comparing their performances. As application, we solve a class of benchmark structural problems modeled by Kirchoff-Love shells for which we consider NURBS geometric parametrizations and we apply the RB method to the solution of this class of fourth order PDEs. We highlight by means of numerical tests, the performance of the RB method applied to standard IGA approximation of parametrized shell geometries.

2015

Using the Green's function to simplify and understand dendrites

Willem Anna Mark Wybo

Neurons are endowed with dendrites: tree-like structures that collect and transform inputs. These arborizations are believed to substantially enhance the computational repertoire of neurons. While it has long been known that dendrites are not iso-potential units, only in the last few decades it was shown experimentally that dendritic branches can transform local inputs in a non-linear fashion. This finding led to the subunit hypothesis, which states that within the dendritic tree, inputs arriving in one branch are transformed non-linearly and independently from what happens in other branches. Recent progress in experimental recording techniques shows that this localized dendritic integration contributes to shaping behavior. While it is generally accepted that the dendritic tree induces multiple subunits, many questions remain unanswered. For instance, it is not known how much separation there needs to be between different branches to be able to function as subunits. Consequently, there is no information on how many subunits can coexist along a dendritic arborization. It is also not known what the input-output relation of these subunits would be, or whether these subunits can be modified by input patterns. As a consequence, assessing the effects of dendrites on the workings of networks of neurons remains mere guesswork. During this work, we choose a theory-driven approach to advance our knowledge about dendrites. Theory can help us understand dendrites by deriving accurate, but conceptually simple models of dendrites that still capture their main computational effects. These models can then be analyzed and fully understood, which in turn teaches us how actual dendrites function computationally. Such simple models typically require less computer operations to simulate than highly detailed dendrite models. Hence, they may also increase the speed of network simulations that incorporate dendrites. The Green's function forms the basis for our theory driven approach. We first explored whether it could be used to reduce the cost of simulating dendrite models. One mathematically interesting finding in this regard is that, because this function is defined on a tree graph, the number of equations can be reduced drastically. Nevertheless, we were forced to conclude that reducing dendrites in this way does not yield new information about the subunit hypothesis. We then focused our attention on another way of decomposing the Green's function. We found that the dendrite model obtained in this way reveals much information on the dendritic subunits. In particular, we found that the occurrence of subunits is well predicted by the ratio of input over transfer impedance in dendrites. This allowed us to estimate the number of subunits that can coexist on dendritic trees. We also found that this ratio can be modified by other inputs, in particular shunting conductances, so that the number of subunits on a dendritic tree can be modified dynamically. We finally were able to show that, due to this dynamical increase of the number of subunits, individual branches that would otherwise respond to inputs as a single unit, could become sensitive to different stimulus features. We believe that this model can be implemented in such a way that it simulates dendrites in a highly efficient manner. Thus, after incorporation in standard neural network simulation software, it can substantially improve the accessibility of dendritic network simulations to modelers.

EPFL2017