Parameter Learning Algorithms for Continuous Model Improvement Using Operational Data

Gaussian mixture has been widely used for data modeling and analysis and the EM algorithm is generally employed for its parameter learning. However, the EM algorithm may be trapped into a local maximum of the likelihood and even leads to a wrong result if the number of components is not appropriately set. Recently, the competitive EM (CEM) algorithm for Gaussian mixtures, a new kind of split-and-merge learning algorithm with certain competitive mechanism on estimated components of the EM algorithm, has been constructed to overcome these drawbacks. In this paper, we construct a new CEM algorithm through the Bayesian Ying-Yang (BYY) harmony stop criterion, instead of the previously used MML criterion. It is demonstrated by the simulation experiments that our proposed CEM algorithm outperforms the original one on both model selection and parameter estimation.KeywordsCompetitive EM (CEM) algorithmGaussian mixtureBayesian Ying-Yang (BYY) harmony learningModel selection

The expectation-maximization (EM) algorithm can compute the maximum-likelihood (ML) or maximum a posterior (MAP) point estimate of the mixture models or latent variable models such as latent Dirichlet allocation (LDA), which has been one of the most popular probabilistic topic modeling methods in the past decade. However, batch EM has high time and space complexities to learn big LDA models from big data streams. In this paper, we present a fast online EM (FOEM) algorithm that infers the topic distribution from the previously unseen documents incrementally with constant memory requirements. Within the stochastic approximation framework, we show that FOEM can converge to the local stationary point of the LDA's likelihood function. By dynamic scheduling for the fast speed and parameter streaming for the low memory usage, FOEM is more efficient for some lifelong topic modeling tasks than the state-of-the-art online LDA algorithms to handle both big data and big models (aka, big topic modeling) on just a PC.

Predicting rock burst hazard with incomplete data using Bayesian networks

The end of the 20th century and the advent of the new millennium have brought in a true merger of sciences for the benefit of mankind. The biggest promise it holds is that of improving the quality of human life by the discovery of newer medicines and better cures for diseases such as cancer and heart disease. Pharmaceutical companies and academic institutions alike have not failed to deliver on part of the promise by bringing out technologies and products that have significantly decreased mortality and morbidity associated with these diseases. An increase in the scale and complexity of the technologies has made it increasingly important to develop intelligent tools to analyze their output, and numerous mathematical and statistical techniques have been explored and exploited to do exactly this. Bayesian networks (BN) and similar graphical models for multivariate analysis are being used for analyzing these data with great success. They have made possible a high resolution insight into disease mechanisms like never before. These insights into the biological processes of health and disease have helped identify the appropriate targets for drug discovery and aided in the process of bringing better drugs faster to the market for patients in need. This chapter briefly explains the application and contribution of Bayesian networks to the drug discovery and development process.

In order to realize robust visual tracking in natural environments, a novel algorithm based on adaptive appearance model is proposed. The model can adapt to changes in object appearance over time. A mixture of three Gaussian distributions models the value of each pixel. An online Expectation Maximization (EM) algorithm is developed to update the parameters of the Gaussians. The observation model in the particle filter is designed based on the adaptive appearance model. Numerous experimental results demonstrate that our proposed algorithm can track objects well under illumination change, large pose variation, and partial or full occlusion.Keywordsvisual trackingadaptive appearance modelExpectation Maximization algorithmparticle filter

Finite mixtures of multivariate skew t (MST) distributions have proven to be useful in modelling heterogeneous data with asymmetric and heavy tail behaviour. Recently, they have been exploited as an effective tool for modelling flow cytometric data. A number of algorithms for the computation of the maximum likelihood (ML) estimates for the model parameters of mixtures of MST distributions have been put forward in recent years. These implementations use various characterizations of the MST distribution, which are similar but not identical. While exact implementation of the expectation-maximization (EM) algorithm can be achieved for `restricted' characterizations of the component skew t-distributions, Monte Carlo (MC) methods have been used to fit the `unrestricted' models. In this paper, we review several recent fitting algorithms for finite mixtures of multivariate skew t-distributions, at the same time clarifying some of the connections between the various existing proposals. In particular, recent results have shown that the EM algorithm can be implemented exactly for faster computation of ML estimates for mixtures with unrestricted MST components. The gain in computational time is effected by noting that the semi-infinite integrals on the E-step of the EM algorithm can be put in the form of moments of the truncated multivariate non-central t-distribution, similar to the restricted case, which subsequently can be expressed in terms of the non-truncated form of the central t-distribution function for which fast algorithms are available. We present comparisons to illustrate the relative performance of the restricted and unrestricted models, and demonstrate the usefulness of the recently proposed methodology for the unrestricted MST mixture, by some applications to three real datasets.

Police patrols are used ubiquitously to deter crimes in urban areas. A distinctive feature of urban crimes is that criminals react opportunistically to patrol officers assignments. Compared to strategic attackers (such as terrorists) with a well-laid out plan, opportunistic criminals are less strategic in planning attacks and more flexible in executing them. In this paper, our goal is to recommend optimal police patrolling strategy against such opportunistic criminals. We first build a game-theoretic model that captures the interaction between officers and opportunistic criminals. However, while different models of adversary behavior have been proposed, their exact form remains uncertain. Rather than simply hypothesizing a model as done in previous work, one key contribution of this paper is to learn the model from real-world criminal activity data. To that end, we represent the criminal behavior and the interaction with the patrol officers as parameters of a Dynamic Bayesian Network (DBN), enabling application of standard algorithms such as EM to learn the parameters. Our second contribution is a sequence of modifications to the DBN representation, that allows for a compact representation of the model resulting in better learning accuracy and increased speed of learning of the EM algorithm when used for the modified DBN. These modifications use marginalization approaches and exploit the structure of this problem. Finally, our third contribution is an iterative learning and planning mechanism that keeps updating the adversary model periodically. We demonstrate the efficiency of our learning algorithm by applying it to a real data set of criminal activity obtained from the police department of University of Southern California (USC) situated in Los Angeles, USA. We project a significant reduction in crime rate using our planning strategy as opposed to the actual strategy deployed by the police department. We also demonstrate the improvement in crime prevention in simulations when we use our iterative planning and learning mechanism compared to just learning once and planing. This work was done in collaboration with the police department of USC.

This paper presents the experimental evaluation of auditory cognition's effects on visual cognition of video. The influences of seven auditory stimuli on visual recognition are investigated based on experimental data of key-down operations. The key-down operations for locating a moving target by visual and auditory images are monitored by an experiment system originally made by devices including VTR, CRT, Data Recorder, etc.. Regression analysis and EM algorithm are applied to analyzing the experiment data of 350 key-down operations, made with 50 people and 7 auditory stimulus types. The following characteristic results about the influence of auditory stimulus on visual recognition are derived. Firstly, seven people responded too early for every experiment. The average of and the standard deviation of their response times are 439[ms] and 231[ms] respectively. Secondly, the other forty three people responded about 10[ms] after at cases, in which auditory images were presented 30[ms] or 60[ms] before visual images. Also they responded about 10[ms] early at the other cases. Thirdly, as the visual image was dominant information used for the key-down decision making, apparent effects of auditory images on the key-down operation were not measured. Averages and standard deviations of distributions measured by EM algorithm, regarding to 7 auditory stimulus types, are considered and verified with the Card's MHP model of human response.

We consider system identification (learning) problems for Gaussian hidden Markov models (GHMMs). We propose an algorithm to tackle the cases where the data is recorded in aggregate (collective) form generated by a large population of individuals following a certain dynamics. Our parameter learning algorithm is built upon the expectation-maximization algorithm with a novel expectation step proposed recently known as the collective Gaussian forward-backward algorithm. The proposed learning algorithm generalizes the traditional Baum-Welch learning algorithm for GHMMs as it naturally reduces to the latter in case of individual observations.

This article introduces a state-space model for the dynamic modeling of curve sequences within the framework of railway switches online monitoring. In this context, each curve has the peculiarity of being subject to multiple changes in regime. The proposed model consists of a specific latent variable regression model whose coefficients are supposed to evolve dynamically in the course of time. Its parameters are recursively estimated across a sequence of curves through an online Expectation-Maximization (EM) algorithm. The experimental study conducted on two real power consumption curve sequences from the French high speed network has shown encouraging results.

With the increasing demand for analytical calculation for datasets of different scales and types, a diversity of clustering algorithms have been developed in recent years, among which the EM clustering algorithm has a good clustering effect and has remained popular in likelihood applications. Based on the superiority of the EM algorithm on clustering small-scale datasets and to further test the performance of the BIRCH tree in classification application, in this chapter, an integration-based fast incremental EM clustering algorithm is presented for multiple-percept detection tasks in image sequences for mobile robotic applications in an unknown environment. Basically, the proposed algorithm first applies the standard EM algorithm to group image-patch-based feature vectors extracted from an image perceptually. Then, a fast tree-based approximate nearest neighbor classifier is employed to integrate the clustering results. Experiments performed in an outdoor environment demonstrate the efficacy of the proposed method.

The traditional Expectation-Maximization (EM) training of Gaussian Mixture Model (GMM) is essentially a batch mode procedure which requires the multiple data samples with the sufficient size to update the model parameters. This severely limits the deployment and adaptation of GMM in many real-time online systems since the newly observed data samples are expected to be incorporated into the system upon available via retraining the model. This paper presents a new online incremental EM training procedure of GMM, which aims to perform the EM training incrementally and so can adapt GMM online sample by sample. The proposed method is extended on two kinds of EM algorithms for GMM, namely, Split-and-Merge EM and the traditional EM. Experiments on both the synthetic data and a speech processing task show the advantages and efficiency of the new method.

This chapter introduces several fundamental models and algorithms for probabilistic clustering, including mixture models, Expectation-Maximization (EM) algorithm, and probabilistic topic models. It focuses on general framework of modeling, the probabilistic explanation, the standard algorithms to learn the model, and its applications. Mixture models are probabilistic models which are increasingly used to find the clusters for univariate and multivariate data. The chapter begins the discussion of mixture models, in which the values of the discrete latent variables can be interpreted as the assignments of data points to specific components of the mixture. It uses the Gaussian mixture model to motivate the EM algorithm in an informal way, and then gives a more general view of the EM algorithm, which is a standard learning algorithm for many probabilistic models. The chapter presents two popular probabilistic topic models, i.e., probabilistic latent semantic analysis (PLSA) and latent Dirichlet allocation (LDA), for document clustering and analysis.

In this paper, a methodology based on Bayesian Network (BN) was proposed to deal with the difficulty of risk analysis in RoPax transport. Based on data collection and expert survey, BN model for RoPax sailing risk analysis was constructed first. Then the Expectation Maximization (EM) algorithm for parameter learning and Evidence Prepropagation Importance Sampling (EPIS) algorithm for reasoning were designed. Finally, a sensitivity analysis was conducted. To validate the model algorithms, a case study on the RoPax system of Bohai gulf in China was provided. Results indicate that the BN model can effectively address the problem of data deficiency and mutual dependency of incidents in risk analysis. It can also model the development process of unexpected hazards and provide decision support for risk mitigation.

This paper presents new methods for the recognition and categorization of object properties such as surface texture, weight, and compliance using a multi-modal artificial skin mounted on both arms of a humanoid. In addition, it introduces two novel feature descriptors, which are useful for providing high-level information to learning algorithms. The artificial skin has built-in 3-axis accelerometer, normal force, proximity, and temperature sensors. To explore different surface textures and weights, objects were left sliding between the NAO humanoid's arms. The caused vibration was detected by accelerometers. Surface texture and weight recognition models were learned from the extracted features of the vibration signals thanks to two learning algorithms, namely the support vector machine (SVM) and the Expectation Maximization (EM). In order to recognize objects having different compliances, SVM and EM took into account total amount of forces applied by the arms to hold the object firmly. The experimental results show that the humanoid can distinguish between different objects having different surface textures and weights with a recognition rate of 100%. Furthermore, it can categorize objects with hard and soft surfaces and classify objects having similar compliance with 100% and 70% accuracy rates respectively.