Qualitatively faithful quantitative prediction

We describe an approach to machine learning from numerical data that combines both qualitative and numerical learning. This approach is carried out in two stages: (1) induction of a qualitative model from numerical examples of the behaviour of a physical system, and (2) induction of a numerical regression function that both respects the qualitative constraints and fits the training data numerically. We call this approach Q² learning, which stands for Qualitatively faithful Quantitative learning. Induced numerical models are “qualitatively faithful” in the sense that they respect qualitative trends in the learning data. Advantages of Q² learning are that the induced qualitative model enables a (possibly causal) explanation of relations among the variables in the modelled system, and that numerical predictions are guaranteed to be qualitatively consistent with the qualitative model which alleviates the interpretation of the predictions. Moreover, as we show experimentally the qualitative model's guidance of the quantitative modelling process leads to predictions that may be considerably more accurate than those obtained by state-of-the-art numerical learning methods. The experiments include an application of Q² learning to the identification of a car wheel suspension system—a complex, industrially relevant mechanical system.     

Models and methods for plan diagnosis

We consider a model-based diagnosis approach to the diagnosis of plans. Here, a plan performed by some agent(s) is considered as a system to be diagnosed. We introduce a simple formal model of plans and plan execution where it is assumed that the execution of a plan can be monitored by making partial observations of plan states. These observed states are used to compare them with states predicted based on (normal) plan execution. Deviations between observed and predicted states can be explained by qualifying some plan steps in the plan as behaving abnormally. A diagnosis is a subset of plan steps qualified as abnormal that can be used to restore the compatibility between the predicted and the observed partial state. Besides minimum and subset minimal diagnoses, we argue that in plan-based diagnosis maximum informative diagnoses should be considered as preferred diagnoses, too. The latter ones are diagnoses that make the strongest predictions with respect to partial states to be observed in the future. We show that in contrast to minimum diagnoses, finding a (subset minimal) maximum informative diagnosis can be achieved in polynomial time. Finally, we show how these diagnoses can be found efficiently if the plan is distributed over a number of agents.     

Constraint-based and SAT-based diagnosis of automotive configuration problems

We compare the concepts and computation of optimized diagnoses in the context of Boolean constraint based knowledge systems of automotive configuration, namely the preferred minimal diagnosis and the minimum weighted diagnosis. In order to restore the consistency of an over-constrained system w.r.t. a strict total order of the user requirements, the preferred minimal diagnosis tries to keep the most preferred user requirements and can be computed, for example, by the FASTDIAG algorithm. In contrast, partial weighted MinUNSAT solvers aim to find a set of unsatisfied clauses with the minimum sum of weights, such that the diagnosis is of minimum weight. It turns out that both concepts have similarities, i.e., both deliver an optimal minimal correction subset. We show use cases from automotive configuration where optimized diagnoses are desired. We point out theoretical commonalities and prove the reducibility of both concepts to each other, i.e., both problems are FP^NP-complete, which was an open question. In addition to exact algorithms we present greedy algorithms. We evaluate the performance of exact and greedy algorithms on problem instances based on real automotive configuration data from three different German car manufacturers, and we compare the time and quality tradeoff.     

Controlling the complexity in model-based diagnosis

We present IDA — an incrementaldiagnosticalgorithm which computes minimal diagnoses from diagnoses, and not from conflicts. As a consequence of this, and by using different models, one can control the computational complexity. In particular, we show that by using a model of the normal behavior, the worst-case complexity of the algorithm to compute thek+1st minimal diagnosis isO(n ^2k ), wheren is the number of components. On the practical side, an experimental evaluation indicates that the algorithm can efficiently diagnose devices consisting of a few thousand components. We propose to use a hierarchy of models: first a structural model to compute all minimal diagnoses, then a normal behavior model to find the additional diagnoses if needed, and only then a fault model for their verification. IDA separates model interpretation from the search for minimal diagnoses in the sense that the model interpreter is replaceable. In particular, we show that in some domains it is advantageous to use the constraint logic programming system CLP(ß) instead of a logic programming system like Prolog.This is an extended version of the paper by Igor Mozeti, A polynomial-time algorithm for model-based diagnosis, which appears in theProc. European Conf. on Artificial Intelligence, ECAI-92, ed. B. Neumann (Wiley, 1992) pp. 729–733.     

Rigorous modeling of gypsum solubility in Na–Ca–Mg–Fe–Al–H–Cl–H2O system at elevated temperatures

Precipitation and scaling of calcium sulfate have been known as major problems facing process industries and oilfield operations. Most scale prediction models are based on aqueous thermodynamics and solubility behavior of salts in aqueous electrolyte solutions. There is yet a huge interest in developing reliable, simple, and accurate solubility prediction models. In this study, a comprehensive model based on least-squares support vector machine (LS-SVM) is presented, which is mainly devoted to calcium sulfate dihydrate (or gypsum) solubility in aqueous solutions of mixed electrolytes covering wide temperature ranges. In this respect, an aggregate of 880 experimental data were gathered from the open literature in order to construct and evaluate the reliability of presented model. Solubility values predicted by LS-SVM model are in well accordance with the observed values yielding a squared correlation coefficient (R ²) of 0.994. Sensitivity of the model for some important parameters is also checked to ascertain whether the learning process has succeeded. At the end, outlier diagnosis was performed using the method of leverage value statistics to find and eliminate the falsely recorded measurements from assembled dataset. Results obtained from this study indicate that LS-SVM model can successfully be applied in predicting accurate solubility of calcium sulfate dihydrate in Na–Ca–Mg–Fe–Al–H–Cl–H₂O system over temperatures ranging from 283.15 to 371.15 K.     

An integrated catchment-coastal modelling system for real-time water quality forecasts

This paper presents a new marine water quality forecasting system for real-time and short-term predictions. The forecasting system comprises an integrated catchment-coastal model and a database management system. The integrated model is validated in an Irish catchment-coastal system using hydrodynamic and water quality data. The forecasting system was then used to provide short-term and real-time forecasts of Escherichia coli (E. coli) and Intestinal Enterococci concentrations (IE) in the near-shore coastal waters of Bray, Ireland. Two hind-cast scenarios were simulated: 5F in which predictions were based on rainfall forecasts only; and I-5F where forecasts of 5F were improved by incorporating real-time data. Results indicate that predictions of E. coli of scenario I-5F are improved. Also predicted IE concentrations by Scenario 5F were comparably higher than the I-5F predications, but due to the wide scatter of observed IE concentrations, the superiority of one scenario over the second could not be definitively determined.     

Mathematical modeling of group product recommendation with partial information: How many ratings do we need?

Product recommendation is one of the most important services in the Internet. In this paper, we consider a product recommendation system which recommends products to a group of users. The recommendation system only has partial preference information on this group of users: a user only indicates his preference to a small subset of products in the form of ratings. This partial preference information makes it a challenge to produce an accurate recommendation. In this work, we explore a number of fundamental questions. What is the desired number of ratings per product so to guarantee an accurate recommendation? What are some effective voting rules in summarizing ratings? How users’ misbehavior such as cheating, in product rating may affect the recommendation accuracy? What are some efficient rating schemes? To answer these questions, we present a formal mathematical model of a group recommendation system. We formally analyze the model. Through this analysis we gain the insight to develop a randomized algorithm which is both computationally efficient and asymptotically accurate in evaluating the recommendation accuracy under a very general setting. We propose a novel and efficient heterogeneous rating scheme which requires equal or less rating workload, but can improve over a homogeneous rating scheme by as much as 30%. We carry out experiments on both synthetic data and real-world data from TripAdvisor. Not only we validate our model, but also we obtain a number of interesting observations, i.e., a small of misbehaving users can decrease the recommendation accuracy remarkably. For TripAdvisor, one hundred ratings per product is sufficient to guarantee a high accuracy recommendation. We believe our model and methodology are important building blocks to refine and improve applications of group recommendation systems.     

Evaluation of streamflow predictions by the IHACRES rainfall-runoff model in two South African catchments

In many South African catchments, water is an increasingly limited and highly fluctuating resource. Accurate prediction of low flows is especially vital if water resource managers are to successfully balance the growing needs of agriculture, industry and rural and urban populations, while maintaining the ecological health of aquatic and riparian ecosystems. Existing hydrological models in use in South Africa suffer from a number of disadvantages. They are complex, over-parameterised, data demanding and expensive to use. IHACRES, a lumped conceptual model requiring minimal input data, is less limited by these problems, and has the potential to advance our understanding of streamflow patterns and predict how these may be altered by land-use change. The purpose of this paper is to evaluate IHACRES performance for two South African catchments: Lambrechtsbos A (a 31 ha research catchment) and Groot-Nylrivier (74 km²). IHACRES predicted streamflow at Lambrechtsbos A with useful accuracy (pre-afforestation period, R²>0.81; bias <26 mm/yr; post-afforestation period, R²=0.81, bias=8.4 mm/yr). With prior knowledge of changes in annual evapotranspiration, predictions of land-use impacts on flow regime may be satisfactorily predicted. Simulations of flows in the Groot-Nylrivier catchment were found to be of useful accuracy for relatively short periods of 2–3 yr, but performance over longer time periods was reduced by poor predictions in certain years. We ascribe this primarily to poor catchment-average rainfall estimation following certain storms in some years. Our simulations highlighted a tendency for IHACRES to underestimate quick flow events, especially at times when the greater part of a catchment is dry. Further model development is required to overcome these problems. IHACRES shows great potential in linking proposed land-use change to altered flow regimes, and efficiently describing the flow characteristics within catchments. However, poor estimation of average rainfall in larger catchments is a limitation that needs to be overcome before long-term flow regimes of non-research catchments may be predicted with confidence.     

Measurement and Prediction of Communication Delays in Myrinet Networks

This paper describes a series of experiments carried out to determine if it is possible to accurately predict the delays of inter-node communication in a PC cluster system interconnected with a Myrinet switch network. Prediction accuracy is affected not only by the software and hardware overhead involved in network communication, but also interference from concurrent message streams. Based on extensive measurements using a 14-node Myrinet cluster system, it is determined that (1) the simple linear model typically used to model communication delay in networks is insufficient and (2) communication delay behavior with n message streams sharing a common link is more complicated than a simple divide-by-n solution. A piecewise-linear model, based on parameters obtained through experiments, is proposed as a more accurate communication delay prediction method when there is no sharing of communication links. However, if two or more message streams share a common link, then the communication delay is more accurately predicted as being one of a set of discrete values.     

An optimal tumor marker group-coupled artificial neural network for diagnosis of lung cancer

BackgroundEpidemiological statistics has shown that there are approximately 1.2 million new cases of lung cancer diagnosed every year and the death rate of these patients is 17.8%. Earlier diagnosis is key to promote the five-year survival rate of these cancer patients. Some tumor markers have been found to be valuable for earlier diagnosis, but a single marker has limitation in its sensitivity and specificity of cancer diagnosis. To improve the efficiency of diagnosis, several distinct tumor marker groups are combined together using a mathematical evaluation model, called artificial neural network (ANN). Lung cancer markers have been identified to include carcinoembryonic antigen, carcinoma antigen 125, neuron specific enolase, β₂-microglobulin, gastrin, soluble interleukin-6 receptor, sialic acid, pseudouridine, nitric oxide, and some metal ions.MethodsThese tumor markers were measured through distinct experimental procedures in 50 patients with lung cancer, 40 patients with benign lung diseases, and 50 cases for a normal control group. The most valuable were selected into an optimal tumor marker group by multiple logistic regression analysis. The optimal marker group-coupled ANN model was employed as an intelligent diagnosis system.ResultsWe have presented evidence that this system is superior to a traditional statistical method, its diagnosis specificity significantly improved from 72.0% to 100.0% and its accuracy increased from 71.4% to 92.8%.ConclusionsThe ANN-based system may provide a rapid and accurate diagnosis tool for lung cancer.     

NICE h : a higher-order explicit numerical scheme for integration of constitutive models in plasticity

The article introduces, as a result of further development of the first-order scheme NICE, a simple and efficient higher-order explicit numerical scheme for the integration of a system of ordinary differential equations which is constrained by an algebraic condition (DAE). The scheme is based on the truncated Taylor expansion of the constraint equation with order h of the scheme being determined by the highest exponent in the truncated Taylor series. The integration scheme thus conceived will be named NICE ^h , considering both principal premises of its construction. In conjunction with a direct solution technique used to solve the boundary value problem, the NICE ^h scheme is very convenient for integrating constitutive models in plasticity. The plasticity models are defined mostly by a system of algebraic and differential equations in which the yield criterion represents the constraint condition. To study the properties of the new integration scheme, which, like the forward-Euler scheme, is characterised by its implementation simplicity due to the explicitness of its formulations, a damage constitutive model (Gurson–Tvergaard–Needleman model) is considered. The general opinion that the implicit backward-Euler scheme is much more accurate than the thus-far known explicit schemes is challenged by the introduction of the NICE ^h scheme. The accuracy of the higher-order explicit scheme in the studied cases is significantly higher than the accuracy of the classical backward-Euler scheme, if we compare them under the condition of a similar CPU time consumption.     

Evolutionary bi-objective optimization of a semi-arid vegetation dynamics model with NDVI and σ satellite data

Satellite radar backscattering coefficient σ⁰ data from ENVISAT-ASAR and Normalized Difference Vegetation Index (NDVI) data from SPOT-VEGETATION are assimilated in the STEP model of vegetation dynamics. The STEP model is coupled with a radiative transfer model of the radar backscattering and NDVI signatures of the soil and herbaceous vegetation. These models are driven by field data (rainfall time series, soil properties, etc.). While some model parameters have fixed values, some other parameters have target values to be optimized. The study focuses on a well documented 1 km² homogeneous area in a semi-arid region (Gourma, Mali).We here investigate whether departures between model predictions and the corresponding data result from field data errors, in situ data lack of representativeness or some model shortcomings. For this purpose we introduce an evolutionary strategy (ES) approach relying on a bi-objective function to be minimized in the data assimilation/inversion process. Several numerical experiments are conducted, in various mono-objective and bi-objective modes, and the performances of the model predictions compared in terms of NDVI, backscattering coefficient, leaf area index (LAI) and biomass.It is shown that the bi-objective ES leads to improved model predictions and also to a better readability of the results by exploring the Pareto front of optimal and admissible solutions. It is also shown that the information brought from the optical sensor and the radar is coherent; that the corresponding radiative transfer models are also coherent; that the representativeness of in situ data can be compared to satellite data through the modeling process. However some systematic biases on the biomass predictions (errors in the range 140 to 300 kg ha^− 1) are observed. Thanks to the bi-objective ES, we are able to identify some likely shortcoming in the vegetation dynamics model relating the LAI to the biomass variables.     

Conditional diagnosability of balanced hypercubes under the MM∗ model

Fault diagnosis plays an important role in ensuring the reliability of a massive multiprocessor system. Diagnosability of a system is the maximum number of faulty nodes guaranteed to be identified during the diagnosis process, and thus is a critical metric to the reliability of the system. To have a greater number of identified faulty nodes, a new measure called conditional diagnosability for fault diagnosis was introduced, which has a normally used assumption. This paper addresses the conditional diagnosability of balanced hypercubes under the MM^? model, which is a realistic model to the fault diagnosis of a system. We show that the conditional diagnosability of the n-dimensional balanced hypercube is 4n?4 for n≥2.     

Three-valued diagnosable systems: Diagnosability,optimal design and fault identification algorithm

A three-valued model for computer system diagnosis is proposed in this paper. A subsystem is regurded as composed of two components, a task processor and a communication processor. Accordingly, subsystem faults are classified as type 1 or type 1/2 faults, representing subsystem faults with communication processors being faulty or those with communication processors being fault-free, The concepts of virtual tests andt ¹ /t ^1/2 fault diagnosis are introduced. Andt ¹ /t ^1/2 fault diagnosis with centralized control and that with distributed control are studied respectively. The problems addressed here include diagnosability, optimal design and fault identification algorithm. The results show clearly thatt ¹ /t ^1/2 fault diagnosis can afford to identify significantly more faults thant-fault diagnosis under the same structural constraint.     

Parameter uncertainty and temporal dynamics of sensitivity for hydrologic models: A hybrid sequential data assimilation and probabilistic collocation method

In this study, a hybrid sequential data assimilation and probabilistic collocation (HSDAPC) approach is proposed for analyzing uncertainty propagation and parameter sensitivity of hydrologic models. In HSDAPC, the posterior probability distributions of model parameters are first estimated through a particle filter method based on streamflow discharge data. A probabilistic collocation method (PCM) is further employed to show uncertainty propagation from model parameters to model outputs. The temporal dynamics of parameter sensitivities are then generated based on the polynomial chaos expansion (PCE) generated by PCM, which can reveal the dominant model components for different catchment conditions. The maximal information coefficient (MIC) is finally employed to characterize the correlation/association between model parameter sensitivity and catchment precipitation, potential evapotranspiration and observed discharge. The proposed method is applied to the Xiangxi River located in the Three Gorges Reservoir area. The results show that: (i) the proposed HSDAPC approach can generate effective 2nd and 3rd PCE models which provide accuracy predictions; (ii) 2^nd-order PCE, which can run nearly ten time faster than the hydrologic model, can capably represent the original hydrological model to show the uncertainty propagation in a hydrologic simulation; (iii) the slow (R_s) and quick flows (R_q) in Hymod show significant sensitivities during the simulation periods but the distribution factor (α) shows a least sensitivity to model performance; (iv) the model parameter sensitivities show significant correlation with the catchment hydro-meteorological conditions, especially during the rainy period with MIC values larger than 0.5. Overall, the results in this paper indicate that uncertainty propagation and temporal sensitivities of parameters can be effectively characterized through the proposed HSDAPC approach.     

Artificial neural networks for rapid WWTP performance evaluation: Methodology and case study

Reliable performance evaluation of wastewater treatment plants (WWTPs) can be done by simulating the plant behavior over a wide range of influent disturbances, including series of rain events with different intensity and duration, seasonal temperature variations, holiday effects, etc. Such simulation-based WWTP performance evaluations are in practice limited by the long simulation time of the mechanistic WWTP models. By moderate simplification (avoiding big losses in prediction accuracy) of the mechanistic WWTP model only a limited reduction of the simulation time can be achieved. The approach proposed in this paper combines an influent disturbance generator with a mechanistic WWTP model for generating a limited sequence of training data (4 months of dynamic data). An artificial neural network (ANN) is then trained on the available WWTP input-output data, and is subsequently used to simulate the remainder of the influent time series (20 years of dynamic data) generated with the influent disturbance generator. It is demonstrated that the ANN reduces simulation time by a factor of 36, even when including the time needed for the generation of training data and for ANN training. For repeated integrated urban wastewater system simulations that do not require repeated training of the ANN, the ANN reduces simulation time by a factor of 1300 compared to the mechanistic model. ANN prediction of effluent ammonium, BOD₅ and total suspended solids was good when compared to mechanistic WWTP model predictions, whereas prediction of effluent COD and total nitrogen concentrations was a bit less satisfactory. With correlation coefficients R² > 0.95 and prediction errors lower than 10%, the accuracy of the ANN is sufficient for applications in simulation-based WWTP design and simulation of integrated urban wastewater systems, especially when taking into account the uncertainties related to mechanistic WWTP modeling.