Use of contexts in language model interpolation and adaptation

Language models (LMs) are often constructed by building multiple individual component models that are combined using context independent interpolation weights. By tuning these weights, using either perplexity or discriminative approaches, it is possible to adapt LMs to a particular task. This paper investigates the use of context dependent weighting in both interpolation and test-time adaptation of language models. Depending on the previous word contexts, a discrete history weighting function is used to adjust the contribution from each component model. As this dramatically increases the number of parameters to estimate, robust weight estimation schemes are required. Several approaches are described in this paper. The first approach is based on MAP estimation where interpolation weights of lower order contexts are used as smoothing priors. The second approach uses training data to ensure robust estimation of LM interpolation weights. This can also serve as a smoothing prior for MAP adaptation. A normalized perplexity metric is proposed to handle the bias of the standard perplexity criterion to corpus size. A range of schemes to combine weight information obtained from training data and test data hypotheses are also proposed to improve robustness during context dependent LM adaptation. In addition, a minimum Bayes' risk (MBR) based discriminative training scheme is also proposed. An efficient weighted finite state transducer (WFST) decoding algorithm for context dependent interpolation is also presented. The proposed technique was evaluated using a state-of-the-art Mandarin Chinese broadcast speech transcription task. Character error rate (CER) reductions up to 7.3 relative were obtained as well as consistent perplexity improvements. © 2012 Elsevier Ltd. All rights reserved.

Generalization in Adaptation to Stable and Unstable Dynamics

Humans skillfully manipulate objects and tools despite the inherent instability. In order to succeed at these tasks, the sensorimotor control system must build an internal representation of both the force and mechanical impedance. As it is not practical to either learn or store motor commands for every possible future action, the sensorimotor control system generalizes a control strategy for a range of movements based on learning performed over a set of movements. Here, we introduce a computational model for this learning and generalization, which specifies how to learn feedforward muscle activity in a function of the state space. Specifically, by incorporating co-activation as a function of error into the feedback command, we are able to derive an algorithm from a gradient descent minimization of motion error and effort, subject to maintaining a stability margin. This algorithm can be used to learn to coordinate any of a variety of motor primitives such as force fields, muscle synergies, physical models or artificial neural networks. This model for human learning and generalization is able to adapt to both stable and unstable dynamics, and provides a controller for generating efficient adaptive motor behavior in robots. Simulation results exhibit predictions consistent with all experiments on learning of novel dynamics requiring adaptation of force and impedance, and enable us to re-examine some of the previous interpretations of experiments on generalization. © 2012 Kadiallah et al.

Risk based lifetime costs assessment of a ground source heat pump (GSHP) system design: Methodology and case study

Space heating accounts for a large portion of the world's carbon dioxide emissions. Ground Source Heat Pumps (GSHPs) are a technology which can reduce carbon emissions from heating and cooling. GSHP system performance is however highly sensitive to deviation from design values of the actual annual energy extraction/rejection rates from/to the ground. In order to prevent failure and/or performance deterioration of GSHP systems it is possible to incorporate a safety factor in the design of the GSHP by over-sizing the ground heat exchanger (GHE). A methodology to evaluate the financial risk involved in over-sizing the GHE is proposed is this paper. A probability based approach is used to evaluate the economic feasibility of a hypothetical full-size GSHP system as compared to four alternative Heating Ventilation and Air Conditioning (HVAC) system configurations. The model of the GSHP system is developed in the TRNSYS energy simulation platform and calibrated with data from an actual hybrid GSHP system installed in the Department of Earth Science, University of Oxford, UK. Results of the analysis show that potential savings from a full-size GSHP system largely depend on projected HVAC system efficiencies and gas and electricity prices. Results of the risk analysis also suggest that a full-size GSHP with auxiliary back up is potentially the most economical system configuration. © 2012 Elsevier Ltd.

Transient flows in enclosures: A generalised approach for modelling the effects of geometry, heat gains and wind

Transient flows in a confined ventilated space induced by a buoyancy source of time-varying strength and an external wind are examined. The space considered has varying cross-sectional area with height. A generalised theoretical model is proposed to investigate the flow dynamics following the activation of an external wind and an internal source of buoyancy. To investigate the effect of geometry, we vary the angle of the wall inclination of a particular geometry in which a point source of constant buoyancy is activated in the absence of wind. Counter-intuitively the ventilation is worse and lower airflow rates are established for geometries of increasing cross-sectional areas with height. We investigate the effect of the source buoyancy strength by comparing two cases: (1) when the buoyancy input is constant and (2) when the buoyancy input gradually increases over time so that after a finite time the total buoyancy inputs for (1) and (2) are identical. The rate at which the source heat gains are introduced has a significant role on the flow behaviour as we find that, in case (2), a warmer layer and a more pronounced overshoot are obtained than in case (1). The effect of assisting and opposing wind on the transient ventilation of an enclosure of constant cross-sectional area with height and constant heat gains is examined. A Froude number Fr is used to define the relative strengths of the buoyancy-induced and wind-induced velocities and five different transient states and their associated critical Fr are identified. © 2010 Elsevier Ltd.

The effect of floor heat source area on the induced airflow in a room

We examine the role of heat source geometry in determining rates of airflow and thermal stratification in natural displacement ventilation flows. We modify existing models to account for heat sources of finite (non-zero) area, such as formed by a sun patch warming the floor of a room. Our model allows for predictions of the steady stratification and ventilation flow rates that develop in a room due to a circular heat source at floor level. We compare our theoretical predictions with predictions for the limiting cases of a point source of heat (yielding a stratified interior), and a uniformly heated floor (yielding a mixed interior). Our theory shows a smooth transition between these two limits, which themselves result in extremes of ventilation, as the ratio of the heat source radius to the room height increases. Our model for the transition from displacement to mixing ventilation is compared to previous work and demonstrates that the transition can occur for smaller sources than previously thought, particularly for rooms with large floor area compared to ceiling height. © 2009 Elsevier Ltd.

Ventilation effectiveness measures based on heat removal: Part 2. Application to natural ventilation flows

The three effectiveness measures based on the ability of a flow to flush buoyancy from a ventilated space proposed by Coffey and Hunt [Ventilation effectiveness measures based on heat removal-part 1. Definitions. Building and Environment, in press, doi:10.1016/j.buildenv.2006.03.016.] are applied to assess and compare two fundamental natural ventilation flows. We focus on the limiting cases of passive displacement and passive mixing ventilation flows during transient conditions. These transient flows occur when, for example, heat is purged from a building at night. Whilst it is widely recognised that mixing flows are less efficient at purging heat than displacement flows, our results indicate that, when a particular zone of a room is considered, displacement ventilation can result in lower effectiveness than mixing ventilation. When a room is considered as a whole, displacement ventilation yields higher effectiveness than mixing ventilation and we quantify these differences in terms of the geometry of the space and opening area. The proposed theoretical predictions are compared with effectiveness deduced from measurements made during laboratory experiments and show good agreement. © 2006 Elsevier Ltd. All rights reserved.

Ventilation effectiveness measures based on heat removal: Part 1. Definitions

The effectiveness of ventilation flows is considered from the perspective of buoyancy (or heat) removal from a space. This perspective is distinct from the standard in which the effectiveness is based on the concentrations of a neutrally buoyant contaminant/passive tracer. Three new measures of effectiveness are proposed based on the ability of a flow to flush buoyancy from a ventilated space. These measures provide estimates of instantaneous and time-averaged effectiveness for the entire space, and local effectiveness at any height of interest. From a generalisation of the latter, a vertical profile of effectiveness is defined. These measures enable quantitative comparisons to be made between different flows and they are applicable when there is a difference in density (as is typical due to temperature differences) between the interior environment and the replacement air. Applications, therefore, include natural ventilation, hybrid ventilation and a range of forced ventilation flows. Finally, we demonstrate how the ventilation effectiveness of a room may be assessed from simple traces of temperature versus time. © 2006 Elsevier Ltd. All rights reserved.

Heat source modelling and natural ventilation efficiency

We compare natural ventilation flows established by a range of heat source distributions at floor level. Both evenly distributed and highly localised line and point source distributions are considered. We demonstrate that modelling the ventilation flow driven by a uniformly distributed heat source is equivalent to the flow driven by a large number of localised sources. A model is developed for the transient flow development in a room with a uniform heat distribution and is compared with existing models for localised buoyancy inputs. For large vent areas the flow driven by localised heat sources reaches a steady state more rapidly than the uniformly distributed case. For small vent areas there is little difference in the transient development times. Our transient model is then extended to consider the time taken to flush a neutrally buoyant pollutant from a naturally ventilated room. Again comparisons are drawn between uniform and localised (point and line) heat source geometries. It is demonstrated that for large vent areas a uniform heat distribution provides the fastest flushing. However, for smaller vent areas, localised heat sources produce the fastest flushing. These results are used to suggest a definition for the term 'natural ventilation efficiency', and a model is developed to estimate this efficiency as a function of the room and heat source geometries. © 2006 Elsevier Ltd. All rights reserved.

LES of heat transfer in electronics

Numerical methods based on the Reynolds Averaged Navier-Stokes (RANS) and Large Eddy Simulation (LES) equations are applied to the thermal prediction of flows representative of those found in and around electronics systems and components. Low Reynolds number flows through a heated ribbed channel, around a heated cube and within a complex electronics system case are investigated using linear and nonlinear LES models, hybrid RANS-LES and RANS-Numerical-LES (RANS-NLES) methods. Flow and heat transfer predictions using these techniques are in good agreement with each other and experimental data for a range of grid resolutions. Using second order central differences, the RANS-NLES method performs well for all simulations. © 2011 Elsevier Inc.

Shroud heat transfer measurements from a rotating cavity with an axial throughflow of air

The paper discusses measurements of heat transfer obtained from the inside surface of the peripheral shroud. The experiments were carried out on a rotating cavity, comprising two 0.985-m-dia disks, separated by an axial gap of 0.065 m and bounded at the circumference by a carbon fiber shroud. Tests were conducted with a heated shroud and either unheated or heated disks. When heated, the disks had the same temperature level and surface temperature distribution. Two different temperature distributions were tested; the surface temperature either increased, or decreased with radius. The effects of disk, shroud, and air temperature levels were also studied. Tests were carried out for the range of axial throughflow rates and speeds: 0.0025 ≤ m ≤ 0.2 kg/s and 12.5 ≤ Ω ≤ 125 rad/s, respectively. Measurements were also made of the temperature of the air inside the cavity. The shroud Nusselt numbers are found to depend on a Grashof number, which is defined using the centripetal acceleration. Providing the correct reference temperature is used, the measured Nusselt numbers also show similarity to those predicted by an established correlation for a horizontal plate in air. The heat transfer from the shroud is only weakly affected by the disk surface temperature distribution and temperature level. The heat transfer from the shroud appears to be affected by the Rossby number. A significant enhancement to the rotationally induced free convection occurs in the regions 2 ≤ Ro ≤ 4 and Ro ≥ 20. The first of these corresponds to a region where vortex breakdown has been observed. In the second region, the Rossby number may be sufficiently large for the central throughflow to affect the shroud heat transfer directly. Heating the shroud does not appear to affect the heat transfer from the disks significantly.

Finite element analysis of compressor disc heat transfer using the transient conduction method with sparse boundary collocation points

Surface temperature measurements from two discs of a gas turbine compressor rig are used as boundary conditions for the transient conduction solution (inverse heat transfer analysis). The disc geometry is complex, and so the finite element method is used. There are often large radial temperature gradients on the discs, and the equations are therefore solved taking into account the dependence of thermal conductivity on temperature. The solution technique also makes use of a multigrid algorithm to reduce the solution time. This is particularly important since a large amount of data must be analyzed to obtain correlations of the heat transfer. The finite element grid is also used for a network analysis to calculate the radiant heat transfer in the cavity formed between the two compressor discs. The work discussed here proved particularly challenging as the disc temperatures were only measured at four different radial locations. Four methods of surface temperature interpolation are examined, together with their effect on the local heat fluxes. It is found that the choice of interpolation method depends on the available number of data points. Bessel interpolation gives the best results for four data points, whereas cubic splines are preferred when there are considerably more data points. The results from the analysis of the compressor rig data show that the heat transfer near the disc inner radius appears to be influenced by the central throughflow. However, for larger radii, the heat transfer from the discs and peripheral shroud is found to be consistent with that of a buoyancy-induced flow.