Bayesian and L1 Approaches to Sparse Unsupervised Learning

The use of L1 regularisation for sparse learning has generated immense research interest, with successful application in such diverse areas as signal acquisition, image coding, genomics and collaborative filtering. While existing work highlights the many advantages of L1 methods, in this paper we find that L1 regularisation often dramatically underperforms in terms of predictive performance when compared with other methods for inferring sparsity. We focus on unsupervised latent variable models, and develop L1 minimising factor models, Bayesian variants of "L1", and Bayesian models with a stronger L0-like sparsity induced through spike-and-slab distributions. These spike-and-slab Bayesian factor models encourage sparsity while accounting for uncertainty in a principled manner and avoiding unnecessary shrinkage of non-zero values. We demonstrate on a number of data sets that in practice spike-and-slab Bayesian methods outperform L1 minimisation, even on a computational budget. We thus highlight the need to re-assess the wide use of L1 methods in sparsity-reliant applications, particularly when we care about generalising to previously unseen data, and provide an alternative that, over many varying conditions, provides improved generalisation performance.

A controller for a miniature intra-aortic ventricular assist device.

Two control algorithms have been developed for a minimally invasive axial-flow ventricular assist device (VAD) for placement in the descending aorta. The purpose of the device is to offload the left ventricle and to augment lower body perfusion in patients with moderate congestive heart failure. The VAD consists of an intra-aortic impeller with a built-in permanent magnet rotor and an extra-aortic stator. The control algorithms, which use pressure readings upstream and downstream of the VAD to determine the pump status, have been tested in a mock circulatory system under two conditions, namely with or without afterload sensitivity. The results give an insight into controller design for an intra-aortic blood pump working in series with the heart.

Initial modelling and controller design for active control of spacecraft microvibrations

Microvibrations, at frequencies between 1 and 1000 Hz, generated by on board equipment, propagate throughout a spacecraft structure affecting the performance of sensitive payloads. The purpose of this work is to investigate strategies to model and reduce these dynamic disturbances by active control. Initial studies were performed by considering a mass loaded panel where the disturbance excitation source consisted of point forces, the objective being to minimise the displacement at an arbitrary output location. Piezoelectric patches acting as sensors and actuators were used. The equations of motion are derived by using Lagrange's equation with modal shapes as Ritz functions. The number of sensors/actuators and their location is variable. The set of equations obtained is then transformed into state variables and some initial controller design studies have been undertaken. These are based on feedback control implemented using a full state feedback and an observer which reconstructs the state vector from the available sensor signal. Here, the basics behind the structural modelling and controller design will be described. This preliminary analysis will also be used to identify short to medium term further work.

Bias-free identification of a linear model-predictive steering controller from measured driver steering behavior

Recent developments in modeling driver steering control with preview are reviewed. While some validation with experimental data has been presented, the rigorous application of formal system identification methods has not yet been attempted. This paper describes a steering controller based on linear model-predictive control. An indirect identification method that minimizes steering angle prediction error is developed. Special attention is given to filtering the prediction error so as to avoid identification bias that arises from the closed-loop operation of the driver-vehicle system. The identification procedure is applied to data collected from 14 test drivers performing double lane change maneuvers in an instrumented vehicle. It is found that the identification procedure successfully finds parameter values for the model that give small prediction errors. The procedure is also able to distinguish between the different steering strategies adopted by the test drivers. © 2006 IEEE.

Implementation of a PI phase angle controller for finite element analysis of the BDFM

Time-stepping finite element analysis of the BDFM for a specific load condition is shown to be a challenging problem because the excitation required cannot be predetermined and the BDFM is not open loops stable for all operating conditions. A simulation approach using feedback control to set the torque and stabilise the BDFM is presented together with implementation details. The performance of the simulation approach is demonstrated with an example and computed results are compared with measurements.

Periodically controlled hybrid systems verifying a controller for an autonomous vehicle

This paper introduces Periodically Controlled Hybrid Automata (PCHA) for describing a class of hybrid control systems. In a PCHA, control actions occur roughly periodically while internal and input actions may occur in the interim changing the discrete-state or the setpoint. Based on periodicity and subtangential conditions, a new sufficient condition for verifying invariance of PCHAs is presented. This technique is used in verifying safety of the planner-controller subsystem of an autonomous ground vehicle, and in deriving geometric properties of planner generated paths that can be followed safely by the controller under environmental uncertainties.

Finite horizon LQR control with limited controller-system communication

We consider finite-horizon LQR control with limited controller-system communication. Within a time-horizon T , the controller can only communicate with the system d

Stabilising terminal cost and terminal controller for ℓasso- MPC: Enhanced optimality and region of attraction

In recent literature, ℓ1-regularised MPC, or ℓasso-MPC, has been recommended for control tasks involving complex requirements on the control signals, for instance, the simultaneous solution of regulation and sharp control allocation for redundantly-actuated systems. This is due to the implicit thresholding ability of LASSO regression. In this paper, a stabilising terminal cost featuring a mixed ℓ1/ℓ2 2 penalty is presented. Then, a candidate terminal controller is computed, with the aim of enlarging the region of attraction. © 2013 EUCA.

Uniform stability of a particle approximation of the optimal filter derivative

Sequential Monte Carlo methods, also known as particle methods, are a widely used set of computational tools for inference in non-linear non-Gaussian state-space models. In many applications it may be necessary to compute the sensitivity, or derivative, of the optimal filter with respect to the static parameters of the state-space model; for instance, in order to obtain maximum likelihood model parameters of interest, or to compute the optimal controller in an optimal control problem. In Poyiadjis et al. [2011] an original particle algorithm to compute the filter derivative was proposed and it was shown using numerical examples that the particle estimate was numerically stable in the sense that it did not deteriorate over time. In this paper we substantiate this claim with a detailed theoretical study. Lp bounds and a central limit theorem for this particle approximation of the filter derivative are presented. It is further shown that under mixing conditions these Lp bounds and the asymptotic variance characterized by the central limit theorem are uniformly bounded with respect to the time index. We demon- strate the performance predicted by theory with several numerical examples. We also use the particle approximation of the filter derivative to perform online maximum likelihood parameter estimation for a stochastic volatility model.

Bayesian learning of noisy Markov decision processes

This work addresses the problem of estimating the optimal value function in a Markov Decision Process from observed state-action pairs. We adopt a Bayesian approach to inference, which allows both the model to be estimated and predictions about actions to be made in a unified framework, providing a principled approach to mimicry of a controller on the basis of observed data. A new Markov chain Monte Carlo (MCMC) sampler is devised for simulation from theposterior distribution over the optimal value function. This step includes a parameter expansion step, which is shown to be essential for good convergence properties of the MCMC sampler. As an illustration, the method is applied to learning a human controller.