Dynamic learning with the EM algorithm for neural networks

In this paper, we derive an EM algorithm for nonlinear state space models. We use it to estimate jointly the neural network weights, the model uncertainty and the noise in the data. In the E-step we apply a forwardbackward Rauch-Tung-Striebel smoother to compute the network weights. For the M-step, we derive expressions to compute the model uncertainty and the measurement noise. We find that the method is intrinsically very powerful, simple and stable.

A new Lagrangian-Eulerian shell-fluid coupling algorithm based on level sets

We propose a computational method for the coupled simulation of a compressible flow interacting with a thin-shell structure undergoing large deformations. An Eulerian finite volume formulation is adopted for the fluid and a Lagrangian formulation based on subdivision finite elements is adopted for the shell response. The coupling between the fluid and the solid response is achieved via a novel approach based on level sets. The basic approach furnishes a general algorithm for coupling Lagrangian shell solvers with Cartesian grid based Eulerian fluid solvers. The efficiency and robustness of the proposed approach is demonstrated with a airbag deployment simulation. It bears emphasis that in the proposed approach the solid and the fluid components as well as their coupled interaction are considered in full detail and modeled with an equivalent level of fidelity without any oversimplifying assumptions or bias towards a particular physical aspect of the problem.

An efficient frequency-domain algorithm for wave scattering problems with application to jet noise

This paper presents a pseudo-time-step method to calculate a (vector) Green function for the adjoint linearised Euler equations as a scattering problem in the frequency domain, for use as a jet-noise propagation prediction tool. A method of selecting the acoustics-related solution in a truncated spatial domain while suppressing any possible shear-layer-type instability is presented. Numerical tests for 3-D axisymmetrical parallel mean flows against semi-analytical reference solutions indicate that the new iterative algorithm is capable of producing accurate solutions with modest computational requirements.

Centrifuge modelling-based verification of response spectra design methods for some vulnerable structures

Residential RC framed structures suffered heavily during the 2001 Bhuj earthquake in Gujarat, India. These types of structures also saw severe damage in other earthquakes such as the 1999 Kocaeli earthquake in Turkey and 921 Ji-Ji earthquake in Taiwan. In this paper the seismic response of residential structures was investigated using physical modelling. Idealised soft storey and top heavy, two degrees of freedom (2DOF) portal frame structures were developed and tested on saturated and dry sand models at 25 g using the Schofield Centre 10-m Beam Centrifuge. It was possible to recreate observed field behaviour using these models. As observed in many of the recent earthquakes, soft storey structures were found to be particularly vulnerable to seismic loads. Elastic response spectra methods are often used in the design of simple portal frame structures. The seismic risk of these structures can be significantly increased due to modifications such as removal of a column or addition of heavy water tanks on the roof. The experimental data from the dynamic centrifuge tests on such soft storey or top-heavy models was used to evaluate the predictions obtained from the response spectra. Response spectra were able to predict seismic response during small to moderate intensity earthquakes, but became inaccurate during strong earthquakes and when soil structure interaction effects became important. Re-evaluation of seismic risk of such modified structures is required and time domain analyses suggested by building codes such as IBC, UBC or NEHRP may be more appropriate. © Springer 2006.

An efficient and robust pitch marking algorithm on the speech waveform for TD-PSOLA

In a Text-to-Speech system based on time-domain techniques that employ pitch-synchronous manipulation of the speech waveforms, one of the most important issues that affect the output quality is the way the analysis points of the speech signal are estimated and the actual points, i.e. the analysis pitchmarks. In this paper we present our methodology for calculating the pitchmarks of a speech waveform, a pitchmark detection algorithm, which after thorough experimentation and in comparison with other algorithms, proves to behave better with our TD-PSOLA-based Text-to-Speech synthesizer (Time- Domain Pitch-Synchronous Overlap Add Text to Speech System).

Natural actor and belief critic: Reinforcement algorithm for learning parameters of dialogue systems modelled as POMDPs

This article presents a novel algorithm for learning parameters in statistical dialogue systems which are modeled as Partially Observable Markov Decision Processes (POMDPs). The three main components of a POMDP dialogue manager are a dialogue model representing dialogue state information; a policy that selects the system's responses based on the inferred state; and a reward function that specifies the desired behavior of the system. Ideally both the model parameters and the policy would be designed to maximize the cumulative reward. However, while there are many techniques available for learning the optimal policy, no good ways of learning the optimal model parameters that scale to real-world dialogue systems have been found yet. The presented algorithm, called the Natural Actor and Belief Critic (NABC), is a policy gradient method that offers a solution to this problem. Based on observed rewards, the algorithm estimates the natural gradient of the expected cumulative reward. The resulting gradient is then used to adapt both the prior distribution of the dialogue model parameters and the policy parameters. In addition, the article presents a variant of the NABC algorithm, called the Natural Belief Critic (NBC), which assumes that the policy is fixed and only the model parameters need to be estimated. The algorithms are evaluated on a spoken dialogue system in the tourist information domain. The experiments show that model parameters estimated to maximize the expected cumulative reward result in significantly improved performance compared to the baseline hand-crafted model parameters. The algorithms are also compared to optimization techniques using plain gradients and state-of-the-art random search algorithms. In all cases, the algorithms based on the natural gradient work significantly better. © 2011 ACM.