A theory of cross-validation error

Abstract

This paper presents a theory of error in cross-validation testing of algorithms for predicting real-valued attributes. The theory justifies the claim that predicting real-valued attributes requires balancing the conflicting demands of simplicity and accuracy. Furthermore, the theory indicates precisely how these conflicting demands must be balanced, in order to minimize cross-validation error. A general theory is presented, then it is developed in detail for linear regression and instance-based learning     

POROUS ELECTRODE THEORY ANALYSIS OF A VARIABLE-DEPTH,FLOW-BY POROUS ELECTRODE

A comparison has been made between a variable-depth and constant-depth, flow-by porous electrode in which a single metal ion deposition occurs in parallel with proton reduction. Porous-electrode theory was applied to calculate the reactor's performance under limiting-current and below limiting-current conditions. The thickness profile of the variable-depth electrode was calculated—as first suggested by Kreysa—by constraining the solution phase potential drop across the electrode to remain constant and independent of streamwise position. For each reactor configuration, identical electrolyte and electrode properties, processing rate, reactant conversion, and inlet thickness were assumed. Under this basis of comparison, the results show: the space-time yield of the constant-depth electrode is larger, but the variable-depth electrode better utilizes the separator's and counterelectrode's surface area; the power requirement of the variable-depth electrode is greater; and, the current efficiency is parameter dependent, but for the majority of the situations likely to be found in wastewater processing, that in the constant-depth electrode is slightly higher. A simple capital-cost analysis shows that the variable-depth electrode may be less expensive if the costs associated with the separator and counterelectrode are greater than those which scale with the electrode volume.     

Advanced Capillary Structures for High Performance Heat Pipes

High performance heat pipes are widely used for the thermal control of electronic devices. Concerning heat transport limitations, typical wick or capillary structures show advantages in some aspects and disadvantages in others. An advanced capillary structure was developed with high thermal effectiveness, low axial pressure drop, high capillary pressure, and a high boiling limit. It combines open minichannels with open microchannels that are manufactured perpendicular on top of the minichannels. The heat transfer coefficient in the evaporator zone, which is a characteristic value for the thermal effectiveness, was up to 3.3 times higher compared to a similar structure without microchannels. A model that combines micro- and macroscopic phenomena was developed. It predicts the heat transfer coefficient with quite good accuracy as long as the microchannels are at least 300 μm.     

Fault detection in linear discrete dynamic systems by a reduced order generalized-likelihood-ratio method

The detectability by conventional step-hypothesized generalized-likelihood-ratio (SHGLR) method for detection of a parameter change (fault detection) in a linear discrete dynamic system is analysed and it is shown that a weakly-diagnosable-space (WDS) exists for dynamics and sensor faults. Based on the fault detectability, a reduced order SHGLR method is then developed which highly improves the detection rate and speed. In the same framework of the GLR method, another reduced order detection scheme is given, which makes the most use of the information about the input and the state of the system to raise the detectability for faults for the case where the step hypothesis cannot be applied effectively.     

Ability of lactic acid bacteria to produce N-heterocycles causing mousy off-flavour in wine

Thirty‐four strains of the genera and species representative of the wine lactic acid bacteria (LAB) were screened in an ethanolic grape juice medium for the production of mousy off‐flavour as determined by an alkaline test strip sensory method. Most Lactobacillus spp., Oenococcus oeni and Leuconostoc mesenteroides produced mousy off‐flavour, whereas Pediococcus spp. were generally incapable of detectable off‐flavour formation. To verify these results, selected LAB were tested for the ability to produce the three sensorially potent N‐heterocycles, 2‐ethyltetrahydropyridine (ETPY), 2‐acetyltetrahydropyridine (ACTPY) and 2‐acetyl‐1‐pyrroline (ACPY), which are associated with the mousy off‐flavour in wine. N‐heterocycle formation was determined by incubating a high density of bacterial cells in a chemically defined N‐heterocycle assay medium, and quantifying the ‘mousy’ compounds by gas chromatography‐mass spectrometry. Most strains produced each of the three N‐heterocycles, with the general order for this capability being Lactobacillus (heterofermentative) > Oenococcus > Pediococcus. Strains of Lactobacillus hilgardii and Lactobacillus brevis were characterised as producing the highest concentration of ACTPY (328–580 μg/L, sensory threshold in water =1.6 μg/L), whereas ACPY and ETPY were produced at a moderate concentration (< 50 and < 10 μg/L, respectively). In addition to the formation of ACPY and ACTPY, three of five strains of the commercially important wine malolactic bacterium, O. oeni, produced the highest concentration of ETPY (87–162 μg/L). Strains of the homofermentative LAB, Pediococcus sp. and Lactobacillus plantarum L11a, however, produced only a relatively low concentration of each mousy N‐heterocycle (≤ 37 μg/L), suggesting that the ability of LAB to produce mousy N‐heterocycles is linked to the pathway of sugar catabolism. The importance of selecting LAB for induction of malolactic fermentation which have limited ability to produce mousy N‐heterocycles is emphasised.     

A non-linear programming approach to the computer-aided design of regulators using a linear-quadratic formulation†

A design technique is proposed for linear regulators in which a feedback controller of fixed structure is chosen to minimize an integral quadratic objective function subject to the satisfaction of integral quadratic constraint functions. Application of a non-linear programming algorithm to this mathematically tractable formulation results in an efficient and useful computer-aided design tool. Particular attention is paid to computational efficiency and various recommendations are made. Two design examples illustrate the flexibility of the approach and highlight the special insight afforded to the designer.     

A NEW APPROACH TO DECENTRALISED PROCESS CONTROL USING PASSIVITY AND SECTOR STABILITY CONDITIONS

A new robust decentralised process control strategy based on the sector stability theorem is proposed in this paper. The interaction between sub-systems is measured by a revised passivity index, which is used in the development of a new stability condition for decentralised control. The controller synthesis method is developed to satisfy the stability condition and H_∞ performance specification. The decentralised control of a boiler furnace is studied as an application example.     

Wine bottle closures: physical characteristics and effect on composition and sensory properties of a Semillon wine 1. Performance up to 20 months post-bottling

A Semillon wine was bottled using 14 different closures: a screw‐cap type, two grades of conventional natural cork, two ‘technical cork’ closures (natural cork with a synthetic component), and 9 closures manufactured from synthetic polymer material. Closure performance was evaluated for physical aspects (e.g. extraction force and energy, change in closure diameter, and ease of closure reinsertion), and for wine composition and sensory properties. Wine under the screw cap closure retained the greatest concentration of sulfur dioxide (SO₂) and ascorbic acid and had the slowest rate of browning. For other closures the trend of SO₂ loss relative to the screw cap closure was apparent from an early stage of testing, and was most evident in the group of synthetic closures, intermediate in the conventional corks, and least evident in the technical cork closures. The loss of SO₂ was in general highly correlated with an increase in wine browning (OD₄₂₀) and the concentration of SO₂ in the wine at six months was a strong predictor of future browning in the wine, particularly after eighteen months. Neither the concentration of dissolved oxygen at bottling (0.6–3.1 mg/L), nor the physical closure measures were predictors of future browning. For several closures upright storage tended to accelerate loss of SO₂ from the wine, but in many cases this effect was marginal. The closures differed widely in regard to physical characteristics, and in general synthetic corks appeared least ‘consumer‐friendly’ in terms of extraction forces, energies, and ease of closure re‐insertion, but there was a trend for natural cork closures to exhibit larger variability in physical characteristics than technical cork and synthetic closures. Sensory analysis indicated large differences in wine flavour properties, with closures which tended to result in the best retention of free SO₂ having wine sensory scores for ‘citrus’ that were generally high whilst scores for the attributes ‘developed’/‘oxidised’ were low. The situation was reversed for wine under closures that performed poorly in the retention of free SO₂. It was found that below a critical level of free SO₂ remaining in the wine, closures exhibited substantially higher ‘oxidised’ aroma. Whilst trichloroanisole‐type (TCA) taint was a noticeable problem for some cork and technical cork closures, any plastic‐type taint appeared not to be a problem with most synthetic closures.     

Effect of protein content, casein–whey protein ratio and pH value on the foaming properties of skimmed milk

The average bubble diameter (ABD) and density of skimmed milk foam decreased with increasing protein content, up to 4%, while drainage was widely comparable. Samples with casein–whey protein ratios (C/W) of 94/6 and 7/93 showed smaller ABD than samples from mixtures with C/W ratios between 60/40 and 20/80. Foams prepared from samples with C/W of 20/80 exhibited the lowest drainage values after 10 min. Increasing pH from 6.4 to 7.0 led to increased overrun and ABD, while drainage was widely comparable, with a minimum value at pH 6.7. The protein composition of skimmed milk and foams prepared from this milk in relation to protein content, C/W ratio and pH value differed only slightly.