Bayesian Network

Microshrinkages are known as probably the most difficult defects to avoid in high-precision foundry. The presence of this failure renders the casting invalid, with the subsequent cost increment. Modelling the foundry process as an expert knowledge cloud allows properly-trained machine learning algorithms to foresee the value of a certain variable, in this case the probability that a microshrinkage appears within a casting. Extending previous research that presented outstanding results with a Bayesian-network-based approach, we have adapted and tested an artificial neural network and the K-nearest neighbour algorithm for the same objective. Finally, we compare the obtained results and show that Bayesian networks are more suitable than the rest of the counterparts for the prediction of microshrinkages.

Mechanical properties are the attributes of a metal to withstand several forces and tensions. Specifically, ultimate tensile strength is the force a material can resist until it breaks. The only way to examine this mechanical property is the employment of destructive inspections that renders the casting invalid with the subsequent cost increment. In a previous work we showed that modelling the foundry process as a probabilistic constellation of interrelated variables allows Bayesian networks to infer causal relationships. In other words, they may guess the value of a variable (for instance, the value of ultimate tensile strength). Against this background, we present here the first ultimate tensile strength prediction system that, upon the basis of a Bayesian network, is able to foresee the values of this property in order to correct it before the casting is made. Further, we have tested the accuracy and error rate of the system with data of a real foundry.

Ultimate tensile strength (UTS) is the force a material can resist until it breaks. The only way to examine this mechanical property is the employment of destructive inspections with the subsequent cost increment. Modelling the foundry process as an expert knowledge cloud allows properly-trained machine-learning algorithms to foresee the value of UTS. Extending previous research that presented outstanding results with a Bayesian-network-based approach , we have adapted an ANN and K-nearest-neighbour algorithm for the same objective . We compare the obtained results and show that artificial neural networks are more suitable than the rest of counterparts for the prediction of UTS.

Probabilistic Relational Models (PRMs) extend Bayesian networks (BNs) with the notion of class of relational databases. Because of their richness, learning them is a difficult task. In this paper, we propose a method that learns a PRM from data using the semantic knowledge of an ontology describing these data in order to make the learning easier. To present our approach, we describe an implementation based on an ontology of transformation processes and compare its performance to that of a method that learns a PRM directly from data. We show that, even with small datasets, our approach of learning a PRM using an ontology is more efficient.

Prima di partire per un lungo viaggio devi portare con te la voglia di non tornare piu'. Irene Grandi from the song "Prima di partire per un lungo viaggio". Molti (maligni!) penseranno che la citazione si riferisce al mio non voler mai piu' tornare in Bicocca ... Ebbene no! Il dottorato e' stato un viaggio e come e' giusto, in quanto viaggio, l'ho percorso "gustandomelo". Il dottorato e' stata un'esperienza molto costruttiva e la crescita personale e di ricercatore che ne ho derivato non e' stata tutta merito mio ma, soprattutto, delle persone che mi hanno indirizzato e accompagnato in questo cammino e che ora mi piacerebbe ringraziare una per una. Prima tra tutte la mia relatrice la professoressa Messina ("La Prof"). Ha saputo consigliarmi sulla strada da prendere e, quando mi perdevo, ha saputo sempre guidarmi verso il raggiungimento dei miei obiettivi. La collaborazione con lei e' stata un'importante occasione di crescita per me. "A braccetto" con La Prof non puo' mancare il professor Archetti. Sempre presente anche se spesso non fisicamente. Tante delle idee presenti in questa tesi sono nate dalla discussione con lui, da sue proposte e dalla sua voglia di "provare". Nel professor Sorrenti ho sempre trovato un confidente e un amico, ma anche una persona sul cui giudizio potevo contare e della cui opinione potevo fidarmi. Grazie! I tre anni di dottorato sono stati tre anni pieni di amici incontrati lungo questo cammino: Dario, Yuri (che ha consegnato la mia domanda di tesi, grazie!) Simone, Sergio, i due Daniele (che poi ho ritrovato a Toronto), Sara (Un amico e' per sempre!), Elena, Daniele (che per primo mi ha spiegato particle filter), Axel (che ora iniziera' questo "viaggio"), Paolo, Davide e tanti altri con i quali ho condiviso tanti momenti di scherzi, di pausa, di chiacchera e anche a volte di ..., grazie a tutti, ragazzi! Ma piu' di tutti devo ringraziare le mie due colleghe e compagne di (s)ventura: Elisabetta e Ilaria. Ne abbiamo passate tante insieme, lavorativamente parlando e non solo, ma i miei ringraziamenti sono soprattutto per questo ultimo anno e mezzo in cui, se pur lontana, mi avete sempre e comunque fatto sentire una di voi. Grazie! Il MIND (ex LIFE) e' stato un bellissimo gruppo nel quale lavorare (e per ognuno degli studenti passati di li' conservo un ricordo particolare) ma c'e' un altro gruppo che mi ha accolta e fatta sentire a casa in quest ultimo periodo: COGS alla UofT. Per farmi capire anche da loro cambio lingua per un attimo. IV At UofT I found a great professor, Craig, who always gave me the right advice about the right thing to do and, more importantly, who introduced me to David, whose ideas and suggestions (during our long meetings) arrived at the right moment, a turning point for this work. But I also found a lot of colleagues and friends always there when I need something. I remember the long lunches with Laurent, that made me feeling like at home, the small talks with Kevin who supported me a lot, the "strange" conversations with Tyler and the support of Mark and Aida. But I also keep in my heart the first encouragement received at UofT by Darius ("You have the application for your work and it is interesting enough!") who is having a baby girl! Ma io fino a qui non sarei arrivata senza l'educazione e l'esempio che ho ricevuto dai miei genitori (e so che ora mia madre iniziera' a piangere): mi hanno sempre dato la possibilita' di perseguire i miei sogni anche se questo ha voluto dire per loro "stringere la cinghia" o vedermi andare dall'altra parte del mondo. Meno male che c'e' sempre stato il mio fratellino a tener loro compagnia, grazie Davide (non solo per questo). Sono certa che la mia nonna conservera' copia di questa tesi nel salotto di casa sua. Grazie nonnina, per i peperoni, i broccoletti ma anche per l'orgoglio con il quale sempre mi guardi. Visto che abbiamo fatto trenta facciamo trentuno, e lasciatemi ringraziare tutti gli amici che ho trovato a Toronto per il loro supporto e la loro vicinanza e gli amici di Milano che mi hanno sempre accolta a braccia aperte ogni volta che sono tornata in Italia e per tutti loro Vera, una cara amica, che probabilmente dovra' stampare questa tesi, grazie! Ho dedicato a lui la prima riga di questa tesi, voglio dedicargli anche l'ultima scritta. Ho trovato in te un amico e un confidente prima che un marito. Sai benissimo che senza di te nulla in questi tre anni sarebbe stato possibile e ti ringrazio per tutte le opportunita', gli sproni, gli aiuti e tutto l'amore che ci hai messo. E, non ultimo, per i sogni che ora perseguiamo insieme ... Cavolo! ora sto piangendo anche io ....

Taking into account relationships between interacting objects can improve the understanding of the dynamic model governing their behaviors. Moreover, maintaining a belief about the ongoing activity while tracking allows online activity recognition and improves the tracking task. We investigate the use of Relational Dynamic Bayesian Networks to represent the relationships for the tasks of multi-target tracking and explicitly consider a discrete variable in the state to represent the activity for online activity recognition. We propose a new transition model that accommodates relations and activities and we extend the Particle Filter algorithm to directly track relations between targets while recognizing ongoing activities.

Introduction

River flow prediction is crucial for water resource management, environmental protection, and flood control, yet the nonlinear nature of hydrological data poses challenges. Previous studies using models like Random Forest and LSTM faced limitations in accuracy and computational complexity. This research introduces the hybrid SNIP-ANN model, combining the novel SNIP algorithm with Artificial Neural Networks, to enhance prediction accuracy. The study focuses on time-series analysis of monthly flows in the Columbia and Niger Rivers, vital freshwater sources. The goal is to improve forecasting precision for better water management and reduced uncertainties.



Materials and Methods

The SNIP-ANN model is an advanced hybrid method for analyzing river discharge time series. In this model, the SNIP algorithm first preprocesses the raw streamflow data; it separates meaningful signal components from background noise by eliminating fluctuations and enhancing significant peaks. Then, an artificial neural network (ANN), inspired by the parallel processing nature of the human brain, learns the nonlinear relationships between past and present discharge values using the processed data. In this study, a three-month lag (Q(t-3)) was selected as the optimal input. By combining SNIP’s strength in feature extraction and ANN’s ability in pattern recognition, the SNIP-ANN model demonstrated high accuracy in prediction. The model was trained and validated using historical monthly discharge data from the Columbia and Niger rivers. Evaluation metrics such as RMSE and the correlation coefficient (R) indicated high precision and low error. The model’s simplicity and use of only discharge data make it suitable for real-time water resource management and applications in other domains like meteorology and economics. Additionally, its lightweight structure and short training time allow for implementation in systems lacking advanced computational facilities. The model’s generalizability also makes it a reliable choice for time series analysis under diverse climatic conditions. Applying this approach in other data-driven fields can improve prediction accuracy while reducing computational costs. Overall, the SNIP-ANN model offers an effective, lightweight, and intelligent solution for predicting streamflow behavior in complex hydrological systems.

Results and Discussion

In this study, in the first stage, the temporal modeling and in the second stage, the performance evaluation of monthly streamflow prediction for the Columbia and Niger rivers during the study period was carried out using the Artificial Neural Network (ANN) model and the hybrid SNIP-enhanced ANN (SNIP-ANN) model. To model the time-dependent behavior of streamflow, three time-lagged discharge values (one-, two-, and three-month delays) were used as input parameters. Among them, the three-month lag (Q(t–3)) was selected as the optimal input based on its stronger correlation with hydrological patterns such as seasonal variability, rainfall fluctuations, and groundwater storage. Seventy percent of the data were used for training and thirty percent for testing the models. Based on the evaluation results from R, RMSE, and other metrics such as the Taylor diagram, violin plot, and observed versus predicted plots, it was found that the SNIP-ANN model significantly outperformed the standard ANN in temporal modeling. For instance, in the Columbia River, the SNIP-ANN model achieved an R value of 0.7503 and an RMSE of 865.23, while the ANN model had an R of 0.3249 and a much higher RMSE of 1325.99. Similarly, for the Niger River, the SNIP-ANN reached an R of 0.9286 and an RMSE of 231.49, surpassing the ANN model, which scored an R of 0.8014 and an RMSE of 487.88. Moreover, the scatter plots revealed that the predicted values in the SNIP-ANN model were more concentrated along the ideal line (y = x), indicating higher prediction accuracy. The violin plots further supported this finding by showing that the SNIP-ANN predictions closely matched the actual distribution of flow values, while the ANN model showed greater deviations. In general, the SNIP-ANN model demonstrated superior capability in capturing nonlinear patterns, reducing prediction error, and improving generalization performance. These results highlight the model’s potential for real-time hydrological forecasting, especially under extreme flow conditions. Future research may focus on integrating ensemble learning strategies or additional preprocessing techniques to further enhance prediction accuracy across diverse hydrological scenarios.



Conclusion

In this study, the hybrid SNIP-ANN model was introduced as an innovative approach for river discharge prediction and compared with the traditional ANN model. The results showed that the hybrid model significantly improved prediction accuracy compared to the ANN model. By using a three-month time delay in the input data, the model was able to effectively simulate seasonal fluctuations and periodic changes in river discharge. The correlation coefficient for the hybrid model increased from 0.3249 to 0.7503 for the Columbia River and from 0.8014 to 0.9286 for the Niger River, while the prediction error (RMSE) decreased from 1325.99 to 865.23 and from 883.487 to 497.231, respectively. This improvement in performance was particularly noticeable in simulating extreme discharge fluctuations and filtering out additional noise in the data. The results indicate that the SNIP-ANN model is an effective tool for river discharge prediction and has high potential for applications in water resource management and flood prediction. This model can be highly beneficial in long-term forecasting and addressing challenges related to climate change and hydrological fluctuations.

Introduction

The rainfall-runoff process, which is affected by various hydrological parameters, is one of the most complex hydrological processes and one of the most basic hydrological topics related to understanding and predicting the processes of runoff production and transfer. It is the outlet point of the watershed. Planning and optimal utilization of runoff is one of the essential issues in watersheds. Therefore, knowing the natural capacity of runoff production and simulating rainfall-runoff is very important. Artificial intelligence and the use of neural network models are one of the methods of rainfall-runoff forecasting. An artificial neural network is a method with the ability to learn, understand, master relationships, and resist errors. Today, artificial intelligence black box methods such as self-constructing and self-learning functions have a wide ability to model and predict complex problems.

Materials and Methods

The purpose of this research is to evaluate the performance of the artificial neural network model for rainfall-runoff simulation in the Saghez sub-basin in Kurdistan province. To carry out this research, 18-year (2001-2018) data received daily from the Meteorological Organization and Saghez Regional Water and Hydrometry Company have been used. In this study, two types of meteorological and hydrometric data were used. The meteorological parameters used include precipitation, evaporation, average temperature, maximum and minimum temperature, and the hydrometric parameter used in this research was only discharge. In the Saghez basin, rainfall-runoff changes have always been considered one of the prominent hydrological indicators. Since the turpentine sub-basin is considered an open basin in terms of its nature, precipitation can be considered a suitable alternative for investigating discharge in the study area of this research. As a result, precipitation is selected as a potential input variable and the adequacy of the remaining variables will be tested separately for the neural network model. In this research, the meteorological parameters used include precipitation, evaporation, average temperature, and maximum and minimum temperature, and the hydrometric parameter used in this research was only Dubai. Finally, to simulate rainfall-runoff using an artificial neural network model, scenarios with different input variables were considered. To evaluate and validate the performance results of the simulated model in different scenarios of this study, using four statistical criteria of correlation coefficient (R), root mean square error (RMSE), mean absolute error (MAE), and Nash-Sutcliffe index (NSE) was done.

Results and Discussion

Six investigated scenarios were randomly selected by combining different inputs. In the first scenario, the input variable includes precipitation and the output variable is discharge. In the second scenario, the input variables include precipitation and evaporation and the output variable is discharge. In the third scenario, the input variables include precipitation and average temperature and the output variable is discharge. In the fourth scenario, the input variables include precipitation and flow variables with a one-day delay and the output variable of flow. In the fifth scenario, the input variables include precipitation, average temperature, maximum and minimum temperature, and the output variable is discharge. In the sixth scenario, the input variables include precipitation, evaporation, average temperature, maximum and minimum temperature, and the output variable of discharge. In all six scenarios, the output variable is the flow rate. Also, in the modeling, 70% of the data for the training section and 30% of the data for the test section were examined. According to the final results, the performance of the artificial neural network model in scenario number four (input variables including rainfall and discharge with a one-day delay) among the six developed scenarios, with correlation coefficient values of 0.92, mean squared error of 6.65, the average absolute error is 2.04 and the Nash-Sutcliffe index is 0.84 in the education section with the values of 0.91, 5.34, 1.57, and respectively 0.82 selected as the best combination in the test section, and in terms of statistical performance indicators, the results of the Nash-Sutcliffe index values in the training and test section were closer to one, which indicates a good match between the observed values and It is simulated. Also, the correlation coefficient specifies the amount of agreement and distribution of observational data with the predicted results, which can be said that the error measurement indicators and data distribution in the training and test section are a favorable result for prediction. The amount of discharge in this scenario shows that it has a much better performance than the rest of the scenarios. Also, in the fourth scenario, changes in the time series of observed discharge values against the simulated values in the training and test phase were investigated in the artificial neural network model. According to this figure, compared to the observed value, the simulated flow rate had good accuracy and an acceptable error value.

Conclusion

The obtained results showed that for the sub-basin of turpentine, the algorithm of the artificial neural network model for simulating rainfall-runoff on a suitable daily scale has obtained suitable and acceptable results. So it can be said that artificial neural network modeling has high accuracy and low error for the study area. Also, artificial intelligence models can be used as a useful tool and a reliable approach for water resource managers.

Hidden Markov Models (HMMs) which fall under the class of latent variable models have received widespread attention in many fields of applications. HMMs were initially developed and applied within the context of speech recognition. The theoretical framework underpinning the formalism of HMMs has also evolved over time and has found an exalted place in the theory of stochastic processes. The three problems HMMs are used to resolve were discussed alongside their solutions in this paper. An application to criminal intelligence in unraveling the culprit in a situation involving theft was also carried out and results obtained indicated that the HMMs approach offered a similar result with that of the well-established Dynamic Programming approach.

The article presents a novel algorithm for merging Bayesian networks generated by different methods, such as expert knowledge and data-driven approaches, while leveraging a symmetry-based approach. The algorithm combines the strengths of each input network to create a more comprehensive and accurate network. Evaluations on traffic accident data from Prague in the Czech Republic and accidents on railway crossings demonstrate superior predictive performance, as measured by prediction error metric. The algorithm identifies and incorporates symmetric nodes into the final network, ensuring consistent representations across different methods. The merged network, incorporating nodes selected from both the expert and algorithm networks, provides a more comprehensive and accurate representation of the relationships among variables in the dataset. Future research could focus on extending the algorithm to deal with cycles and improving the handling of conditional probability tables. Overall, t...

Three dimensional gait analysis (3DGA) is a vital element in the multidisciplinary treatment of children with cerebral palsy. As 3DGA generates an extensive amount of data, preprocessing steps are needed to facilitate the clinical interpretation and decision-making process. Pre-processing often includes classification of gait into distinct gait patterns, using either qualitative or quantitative approaches. While the former can be subjective and have limited repeatability, the latter often end up with clinically irrelevant gait patterns . Moreover, so far, both qualitative and quantitative classification methods force CP gait to fit into one single gait pattern category. As CP gait is often a mix of several distinct gait patterns, classifying it into just one pattern may have a limited clinical meaning. Therefore, new strategies for efficient data-analysis need to be explored. This study introduces Bayesian networks (BN) as a new approach for classification of 3DGA data of the ankle and knee in children with CP. The developed BN integrated prior knowledge (clinical expertise) to probabilistically classify CP ankle and knee movements on a continuum of patterns. A group of 139 patients with CP were recruited and divided into a training set (n= 80% of all patients) and a validation set (n= 20% of all patients). The proposed BN reached accuracies between 73% -100% and was found to be successful for classifying ankle and knee joint motion on a continuum of different clinically relevant joint movement patterns.

Constraint-based causal structure learning algorithms rely on the faithfulness pro- perty. For faithfulness, all conditional in- dependencies should come from the sys- tem's causal structure. The problem is that even for linear Gaussian models the pro- perty is not tenable. In this paper, we identify 4 non-causal properties that gene- rate conditional independencies and inves- tigate whether they can be recognized by patterns in the conditional independencies.

The principle of Kolmogorov Minimal Sufficient Statistic (KMSS) states that a model should capture all regularities of the data. The conditional independencies following from the causal structure of the system are the regularities incorporated in a graphical causal model. We prove that for joint probability distributions, the KMSS is described by the Directed Acyclic Graph (DAG) of the minimal Bayesian network if this results in an incompressible description. We prove that a Bayesian network that is the KMSS is faithful. In that case it can be learned from observations and modularity is the most plausible hypothesis. From modularity follows the ability to predict the effect of interventions. On the other hand, if the minimal Bayesian network is compressible, and thus not the KMSS, the above implications cannot be guaranteed. When the non-minimality of the description is due to the compressibility of an individual Conditional Probability Distribution (CPD), the true causal model is an element of the set of minimal Bayesian networks and modularity is still plausible. Faithfulness cannot be guaranteed though. When the concatenation of the descriptions of the CPDs is compressible, the true causal model is not necessarily an element of the set of minimal Bayesian networks. Also modularity may become implausible. This suggests that either there is a kind of meta-mechanism governing some of the mechanisms or either a single mechanism responsible for setting the state of multiple variables.

The author has granted a nonexclusive license allowing Library and Archives Canada to reproduce, publish, archive, preserve, conserve, communicate to the public by telecommunication or on the Internet, loan, distrbute and sell theses worldwide, for commercial or noncommercial purposes, in microform, paper, electronic and/or any other formats.

The R package abn is designed to fit additive Bayesian models to observational datasets. It contains routines to score Bayesian networks based on Bayesian or information theoretic formulations of generalized linear models. It is equipped with exact search and greedy search algorithms to select the best network. It supports a possible blend of continuous, discrete and count data and input of prior knowledge at a structural level. The Bayesian implementation supports random effects to control for one-layer clustering. In this paper, we give an overview of the methodology and illustrate the package's functionalities using a veterinary dataset about respiratory diseases in commercial swine production.

Additive Bayesian networks are types of graphical models that extend the usual Bayesian generalized linear model to multiple dependent variables through the factorisation of the joint probability distribution of the underlying variables. When fitting an ABN model, the choice of the prior of the parameters is of crucial importance. If an inadequate prior -like a too weakly informative one -is used, data separation and data sparsity lead to issues in the model selection process. In this work a simulation study between two weakly and a strongly informative priors is presented. As weakly informative prior we use a zero mean Gaussian prior with a large variance, currently implemented in the Rpackage abn. The second prior belongs to the Student's t-distribution, specifically designed for logistic regressions and, finally, the strongly informative prior is again Gaussian with mean equal to true parameter value and a small variance. We compare the impact of these priors on the accuracy of the learned additive Bayesian network in function of different parameters. We create a simulation study to illustrate Lindley's paradox based on the prior choice. We then conclude by highlighting the good performance of the informative Student's t-prior and the limited impact of the Lindley's paradox. Finally, suggestions for further developments are provided.

The R package abn is designed to fit additive Bayesian network models to observational datasets and contains routines to score Bayesian networks based on Bayesian or information theoretic formulations of generalized linear models. It is equipped with exact search and greedy search algorithms to select the best network, and supports continuous, discrete and count data in the same model and input of prior knowledge at a structural level. The Bayesian implementation supports random effects to control for one-layer clustering. In this paper, we give an overview of the methodology and illustrate the package's functionality using a veterinary dataset concerned with respiratory diseases in commercial swine production.

The R package abn is designed to fit additive Bayesian models to observational datasets. It contains routines to score Bayesian networks based on Bayesian or information theoretic formulations of generalized linear models. It is equipped with exact search and greedy search algorithms to select the best network. It supports a possible blend of continuous, discrete and count data and input of prior knowledge at a structural level. The Bayesian implementation supports random effects to control for one-layer clustering. In this paper, we give an overview of the methodology and illustrate the package's functionalities using a veterinary dataset about respiratory diseases in commercial swine production.

Additive Bayesian networks are types of graphical models that extend the usual Bayesian generalized linear model to multiple dependent variables through the factorisation of the joint probability distribution of the underlying variables. When fitting an ABN model, the choice of the prior of the parameters is of crucial importance. If an inadequate prior -like a too weakly informative one -is used, data separation and data sparsity lead to issues in the model selection process. In this work a simulation study between two weakly and a strongly informative priors is presented. As weakly informative prior we use a zero mean Gaussian prior with a large variance, currently implemented in the Rpackage abn. The second prior belongs to the Student's t-distribution, specifically designed for logistic regressions and, finally, the strongly informative prior is again Gaussian with mean equal to true parameter value and a small variance. We compare the impact of these priors on the accuracy of the learned additive Bayesian network in function of different parameters. We create a simulation study to illustrate Lindley's paradox based on the prior choice. We then conclude by highlighting the good performance of the informative Student's t-prior and the limited impact of the Lindley's paradox. Finally, suggestions for further developments are provided.

It is well known that computing relative approximations of weighted counting queries such as the probability of evidence in a Bayesian network, the partition function of a Markov network, and the number of solutions of a constraint satisfaction problem is NP-hard. In this paper, we settle therefore on an easier problem of computing highconfidence lower bounds and propose an algorithm based on importance sampling and Markov inequality for it. However, a straight-forward application of Markov inequality often yields poor lower bounds because it uses only one sample. We therefore propose several new schemes that extend it to multiple samples. Empirically, we show that our new schemes are quite powerful, often yielding substantially higher (better) lower bounds than state-of-the-art schemes.

In this paper, we consider Hybrid Mixed Networks (HMN) which are Hybrid Bayesian Networks that allow discrete deterministic information to be modeled explicitly in the form of constraints. We present two approximate inference algorithms for HMNs that integrate and adjust well known algorithmic principles such as Generalized Belief Propagation, Rao-Blackwellised Importance Sampling and Constraint Propagation to address the complexity of modeling and reasoning in HMNs. We demonstrate the performance of our approximate inference algorithms on randomly generated HMNs.

This paper extends previously proposed bound propagation algorithm for computing lower and upper bounds on posterior marginals in Bayesian networks. We improve the bound propagation scheme by taking advantage of the directionality in Bayesian networks and applying the notion of relevant subnetwork. We also propose an approximation scheme for the linear optimization subproblems. We demonstrate empirically that while the resulting bounds loose some precision, we achieve 10-100 times speedup compared to original bound propagation using a simplex solver.

Computing the partition function is a key inference task in many graphical models. In this paper, we propose a dynamic importance sampling scheme that provides anytime finite-sample bounds for the partition function. Our algorithm balances the advantages of the three major inference strategies, heuristic search, variational bounds, and Monte Carlo methods, blending sampling with search to refine a variationally defined proposal. Our algorithm combines and generalizes recent work on anytime search [16] and probabilistic bounds of the partition function. By using an intelligently chosen weighted average over the samples, we construct an unbiased estimator of the partition function with strong finite-sample confidence intervals that inherit both the rapid early improvement rate of sampling and the long-term benefits of an improved proposal from search. This gives significantly improved anytime behavior, and more flexible trade-offs between memory, time, and solution quality. We demonstrate the effectiveness of our approach empirically on real-world problem instances taken from recent UAI competitions.

This paper develops a measure for bounding the performance of AND/OR search algorithms for solving a variety of queries over graphical models. We show how drawing a connection to the recent notion of hypertree decompositions allows to exploit determinism in the problem specification and produce tighter bounds. We demonstrate on a variety of practical problem instances that we are often able to improve upon existing bounds by several orders of magnitude.

The paper describes a branch and bound scheme that uses heuristics generated mechanically by the mini-bucket approximation. This scheme is presented and evaluated for optimization tasks such as finding the Most Probable Explanation (MPE ) in Bayesian networks. The mini-bucket scheme yields monotonic heuristics of varying strengths which cause different amounts of pruning, allowing a controlled tradeoff between preprocessing and search. The resulting Branch and Bound with Mini-Bucket heuristic (BBMB), is evaluated using random networks, probabilistic decoding and medical diagnosis networks. Results show that the BBMB scheme overcomes the memory explosion of bucket-elimination allowing a gradual tradeoff of space for time, and of time for accuracy.

The focus of this paper is on the analysis of the cyber security resilience of digital infrastructures deployed by power grids, internationally recognized as a priority since several recent cyber attacks targeted energy systems and in particular the power service. In response to the regulatory framework, this paper presents an analysis approach based on the Bayesian Networks formalism and on real world threat scenarios. Our approach enables analyses oriented to planning of security measures and monitoring, and to forecasting of adversarial behaviours.

Particle filtering algorithms can be used for the monitoring of dynamic systems with continuous state variables and without any constraints on the form of the probability distributions. The dimensionality of the problem remains a limitation of these approaches due to the growing number of particles required for the exploration of the state space. Computer vision problems such as 3D motion tracking are an example of complex monitoring problems which have a high dimensional state space and observation functions with high computational cost. In this article we focus on reducing the required number of particles in the case of monitoring tasks where the state vector and the observation function can be factored. We introduce a particle filtering algorithm based on the Dynamic Bayesian Network (DBN) formalism which takes advantage of a factored representation of the state space for efficiently weighting and selecting the particles. We illustrate the approach on a simulated and a realworld 3D motion tracking tasks.

S. Castagnos† castagno@loria.fr A. Boyer† boyer@loria.fr F. Charpillet† charp@loria.fr ... † Laboratoire Lorrain de Recherche en Informatique et Applications Campus Scientifique, BP 239 54506 Vandœuvre-les-Nancy Cedex ... Résumé : Le terme de filtrage collaboratif [7] ...

Constructing a sensibly functional gene interaction network is highly appealing for better understanding system-level biological processes governing various Populus traits. Bayesian Network (BN) learning provides an elegant and compact statistical approach for modeling causal gene-gene relationships in microarray data. Therefore, it could come with the illumination of functional molecular playing in Biology Systems. In the present study, different forms of gene Bayesian networks were detected on Populus cellular transcriptome data. Markov blankets would likely be emerging at every possible gene regulatory Bayesian network level. Results showed that PtpAffx.1257.4.S1_a_at,1.0 hypothetical protein is the most important in its possible regulatory program. This paper illustrates that the gene network regulatory inference is possible to encapsulate within a single BN model. Therefore, such a BN model can serve as a promising training tool for Populus gene expression data for better future experimental scenarios.

Traffic routes through a street network contain patterns and are no random walks. Such patterns exist for instance along streets or between neighbouring street segments. The extraction of these patterns is a challenging task due to the enormous size of city street networks, the large number of required training data and the unknown distribution of the latter. We apply Bayesian Networks to model the correlations between the locations in space-time trajectories and address the following tasks. We introduce and examine a Bayesian Network Learning algorithm enabling us to handle the complexity and performance requirements of the spatial context. Furthermore, we apply our method to German cities, evaluate the accuracy and analyse the runtime behaviour for different parameter settings.

We introduce a conditional compression problem and propose a fast framework for tackling it. The problem is how to quickly compress a pretrained large neural network into optimal smaller networks given target contexts, e.g. a context involving only a subset of classes or a context where only limited compute resource is available. To solve this, we propose an efficient Bayesian framework to compress a given large network into much smaller size tailored to meet each contextual requirement. We employ a hypernetwork to parameterize the posterior distribution of weights given conditional inputs and minimize a variational objective of this Bayesian neural network. To further reduce the network sizes, we propose a new input-output group sparsity factorization of weights to encourage more sparseness in the generated weights. Our methods can quickly generate compressed networks with significantly smaller sizes than baseline methods.

The complexity and spatial heterogeneity of ecosystem processes driving ecosystem service delivery require spatially explicit models that take into account the different parameters affecting those processes. Current attempts to model ecosystem service delivery on a broad, regional scale often depend on indicator-based approaches that are generally not able to fully capture the complexity of ecosystem processes. Moreover, they do not allow quantification of uncertainty on their predictions. In this paper, we discuss a QGIS plug-in which promotes the use of Bayesian belief networks for regional modelling and mapping of ecosystem service delivery and associated uncertainties. Different types of specific Bayesian belief network output maps, delivered by the plug-in, are discussed and their decision support capacities are evaluated. This plug-in, used in combination with firmly developed Bayesian belief networks, has the potential to add value to current spatial ecosystem service accounting methods. The plug-in can also be used in other research domains dealing with spatial data and uncertainty.

The development of effective techniques for knowledge representation and reasoning (KRR) is a crucial aspect of successful intelligent systems. Different representation paradigms, as well as their use in dedicated reasoning systems, have been extensively studied in the past. Nevertheless, new challenges, problems, and issues have emerged in the context of knowledge representation in Artificial Intelligence (AI), involving the logical manipulation of increasingly large information sets (see for example Semantic Web, BioInformatics and so on). Improvements in storage capacity and performance of computing infrastructure have also affected the nature of KRR systems, shifting their focus towards representational power and execution performance. Therefore, KRR research is faced with a challenge of developing knowledge representation structures optimized for large scale reasoning. This new generation of KRR systems includes graph-based knowledge representation formalisms such as Bayesian Networks (BNs), Semantic Networks (SNs), Conceptual Graphs (CGs), Formal Concept Analysis (FCA), CPnets, GAI-nets, all of which have been successfully used in a number of applications. The goal of this workshop is to bring together the researchers involved in the development and application of graph-based knowledge representation formalisms and reasoning techniques.

Background: Leprosy reactions (LR) are severe episodes of intense activation of the host inflammatory response, of uncertain etiology, today the leading cause of permanent nerve damage in leprosy patients. Several genetic and non-genetic risk factors for LR have been described; however, there are limited attempts to combine this information in order to estimate the risk of a leprosy patient to develop LR. Here we present an artificial intelligence (AI)-based system able to estimate risk of LR using clinical, demographic and genetic data.Methods: The study includes four datasets from different regions of Brazil, totalizing 1,450 leprosy patients followed prospectively for at least two years to assess the occurrence of LR. Data mining using WEKA software was performed following a two-step protocol to select the variables included in the AI system, based on Bayesian Networks and developed using the NETICA software.Results: Analysis of the complete database resulted in a system able to ...

In this research, we dene a cooperative multiagent system where the agents use locally designed Bayesian networks to represent their knowledge. Agents communicate via message passing where the messages are beliefs in shared variables that are represented as probability distributions. Messages are treated as soft evidence in the receiver agents, where the belief in the receiving agent is replaced by the publishing agent’s belief. We call this the oracular assumption, where one agent is an expert or more knowledgeable of particular variables. As a result, the agents are organized in a publishersubscriber hierarchy. We compare and contrast our system with the MSBN multiagent model. Finally, we implement a multiagent system for experimentation using our multiagent system and MSBNs. We devise performance measures to compare the two systems. From this comparison, we provide guidance for the design of probabilistic multiagent systems.

In this paper we analyze the performance of three algorithms for soft evidential update, in which a probability distribution represented by a Bayesian network is modified to a new distribution constrained by given marginals, and closest to the original distribution according to cross entropy. The first algorithm is a new and improved version of the big clique algorithm [1] that utilizes lazy propagation . The second and third algorithm [3] are wrapper methods that convert soft evidence to virtual evidence, in which the evidence for a variable consists of a likelihood ratio. Virtual evidential update is supported in existing Bayesian inference engines, such as Hugin. To evaluate the three algorithms, we implemented BRUSE (Bayesian Reasoning Using Soft Evidence), a new Bayesian inference engine, and instrumented it. The resulting statistics are presented and discussed.

We define a cooperative multiagent system where the agents use locally designed Bayesian networks to represent their knowledge. Agents in the newly defined AEBN model communicate via message passing where the messages are beliefs in shared variables that are represented as probability distributions. Messages are treated as soft evidence in the receiver agents, where the belief in the receiving agent is replaced by the publishing agent's belief. We call this the oracular assumption, where one agent is an expert or more knowledgeable of particular variables. As a result, the agents are organized in a publisher-subscriber hierarchy. A central problem of message passing in probabilistic systems is the so called rumor problem, where cycles in message passing cause redundant influence of beliefs. In this paper, algorithms to identify and solve the rumor problem in the context of our multiagent system are presented. The AEBN model is compared and contrasted with the MSBN multiagent model. We implement several multiagent systems for experimentation using our AEBN multiagent system and the MSBN system and we compare them using several performance measures. From this comparison, we provide guidance for the design of probabilistic multiagent systems.

We present a multiagent organization for data interpretation and fusion in which each agent uses an encapsulated Bayesian network for knowledge representation, and agents communicate by exchanging beliefs (marginal posterior probabilities) on shared variables. We call this organization an Agent-Encapsulated Bayesian Network (AEBN) system. Communication of probabilities among agents leads to rumors, i.e. potential double counting of information. We propose a new and correct method to compensate for rumors in AEBN systems by passing extended messages that contain joint probabilities.

Structural learning of directed acyclic graphs (DAGs) or Bayesian networks has been studied extensively under the assumption that data are independent. We propose a new Gaussian DAG model for dependent data which assumes the observations are correlated according to an undirected network. Under this model, we develop a method to estimate the DAG structure given a topological ordering of the nodes. The proposed method jointly estimates the Bayesian network and the correlations among observations by optimizing a scoring function based on penalized likelihood. We show that under some mild conditions, the proposed method produces consistent estimators after one iteration. Extensive numerical experiments also demonstrate that by jointly estimating the DAG structure and the sample correlation, our method achieves much higher accuracy in structure learning. When the node ordering is unknown, through experiments on synthetic and real data, we show that our algorithm can be used to estimate the correlations between samples, with which we can de-correlate the dependent data to significantly improve the performance of classical DAG learning methods.

Distance education has grown in importance with the advent of the internet. An adequate evaluation of students in this mode is still difficult. Distance tests or occasional on-site exams do not meet the needs of evaluation of the learning process for distance education. Bayesian networks are adequate for simulating several aspects of clinical reasoning. The possibility of integrating them in distance education student evaluation has not yet been explored much. The present work describes a Simulator based on probabilistic networks built to represent knowledge of clinical practice guidelines in Family and Community Medicine. The Bayesian Network, the basis of the simulator, was modeled to playable by the student, to give immediate feedback according to pedagogical strategies adapted to the student according to past performance, and to give a broad evaluation of performance at the end of the game. Simulators structured by Bayesian Networks may become alternatives in the evaluation of s...

Human Immunodeficiency Virus-1 (HIV-1) antiviral resistance is a major cause of antiviral therapy failure and compromises future treatment options. As a consequence, resistance testing is the standard of care. Because of the high degree of HIV-1 natural variation and complex interactions, the role of resistance mutations is in many cases insufficiently understood. We applied a probabilistic model, Bayesian networks, to analyze direct influences between protein residues and exposure to treatment in clinical HIV-1 protease sequences from diverse subtypes. We can determine the specific role of many resistance mutations against the protease inhibitor nelfinavir, and determine relationships between resistance mutations and polymorphisms. We can show for example that in addition to the well-known major mutations 90M and 30N for nelfinavir resistance, 88S should not be treated as 88D but instead considered as a major mutation and explain the subtype-dependent prevalence of the 30N resistan...

Bayesian networks have been used as a mechanism to represent the joint distribution of multiple random variables in a flexible yet interpretable manner. One major challenge in learning the structure of a Bayesian network is how to model networks which include a mixture of continuous and discrete random variables, known as hybrid Bayesian networks. This paper overviews the literature on approaches to handle hybrid Bayesian networks. Typically one of two approaches is taken: either the data are considered to have a joint distribution which is designed for a mixture of discrete and continuous variables, or continuous random variables are discretized, resulting in discrete Bayesian networks. In this paper, we propose a strategy to model all random variables as Gaussian, referred to it as Run it As Gaussian (RAG). We demonstrate that RAG results in more reliable estimates of graph structures theoretically and by simulation studies, than converting continuous random variables to discrete. Both strategies are also implemented on a childhood obesity data set. The two different strategies give rise to significant differences in the optimal graph structures, with the results of the simulation study suggesting that our strategy is more reliable.

Risk models using fault and event trees can be extended with explicit factors, which are states of the system, its users or its environment that influence event probabilities. The factors act as parameters in the risk model, enabling the model to be re-used and also providing a new way to estimate the overall risk of a system with many instances of the risk. A risk model with parameters can also be clearer.

Event trees are a popular technique for modelling accidents in system safety analyses. Bayesian networks are a probabilistic modelling technique representing influences between uncertain variables. Although popular in expert systems, Bayesian networks are not used widely for safety. Using a train derailment case study, we show how an event tree can be viewed as a Bayesian network, making it clearer when one event affects a later one. Since this effect needs to be understood to construct an event tree correctly, we argue that the two notations should be used together. We then show how the Bayesian Network enables the factors that influence the outcome of events to be represented explicitly. In the case study, this allowed the train derailment model to be generalised and applied in more circumstances. Although the resulting model is no longer just an event tree, the familiar event tree notation remains useful.

We describe a method of modelling organisational causes of accidents, using Bayesian Networks. A rigorous method is used to relate interactions within the organisation operating the system to causal factors for accidents. Using examples from a model of the causes of SPAD incidents in the UK railways, we describe how such a model can be used for risk assessment.