|
|
GA 19-07635S: Outputs and Results
Rehák, Branislav
This manuscript aims to deliver a survey of results obtained during the solution of the project No. GA19-07635S of the Czech Science Foundation. The timespan dedicated to the work on this project was 1.3.2019 - 30.6.2022. The main area dealt with were\nnonlinear multi-agent systems and their synchronization, further, attention was paid to some auxiliary results in the area of nonlinear observers. This Report briefly introduces the Project, provides a summary of the results obtained and also sketches an outline how these results will be applied and extended in future.
|
|
|
Diffusion Kalman filtering under unknown process and measurement noise covariance matrices
Vlk, T. ; Dedecius, Kamil
The state-of-the-art algorithms for Kalman filtering in agent networks with information diffusion impose the requirement of well-defined state-space models. In particular, they assume that both the process and measurement noise covariance matrices are known and properly set. This is a relatively strong assumption in the signal processing domain. By design, the Kalman filters are rather sensitive to its violation, which may potentially lead to their divergence. In this paper, we propose a novel distributed filtering algorithm with increased robustness under unknown process and measurement noise covariance matrices. It is formulated as a Bayesian variational message passing procedure for simultaneous analytically tractable inference of states and measurement noise covariance matrices.
|
|
|
Recursive mixture estimation with univariate multimodal Poisson variable
Uglickich, Evženie ; Nagy, Ivan
Analysis of count variables described by the Poisson distribution is required in many application fields. Examples of the count variables observed per a time unit can be, e.g., number of customers, passengers, road accidents, Internet traffic packet arrivals, bankruptcies, virus attacks, etc. If the behavior of such a variable exhibits a multimodal character, the problem of clustering and classification of incoming count data arises. This issue can touch, for instance, detecting clusters of the different behavior of drivers in traffic flow analysis as well as cyclists or pedestrians. This work focuses on the model-based clustering of Poisson-distributed count data with the help of the recursive Bayesian estimation of the mixture of Poisson components. The aim of the work is to explain the methodology in details with an illustrative simple example, so that the work is limited to the univariate case and static pointer.
|
|
|
Ockham's Razor from a Fully Probabilistic Design Perspective
Hoffmann, A. ; Quinn, Anthony
This research report investigates an approach to the design of an Ockham prior penalising parametric complexity in the Hierarchical Fully Probabilistic Design (HFPD) [1] setting. We identify a term which penalises the introduction of an additional parameter in the Wold decomposition. We also derive the objective Ockham Parameter Prior (OPI) in this context, based on earlier work [2], and we show that the two are, in fact, closely related. This confers validity on the HFPD Ockham term.
|
|
|
Transferring Improved Local Kernel Design in Multi-Source Bayesian Transfer Learning, with an application in Air Pollution Monitoring in India
Nugent, Sh. ; Quinn, Anthony
Existing frameworks for multi-task learning [1],[2] often rely on completely modelled relationships between tasks, which may not be available. Recent work [3], [4] has been undertaken on approaches to fully probabilistic methods for transfer learning between two Gaussian Process (GP) tasks. There, the target algorithm accepts source knowledge in the form of a probabilistic prior from a source algorithm, without requiring the target to model their interaction with the source. These strategies have offered robust improvements on current state of the art algorithms, such as the Intrinsic Coregionalization Model. The Bayesian Transfer Learning algorithm proposed in [4], was found to provide robust, positive\ntransfer. This algorithm was then extended to accommodate knowledge transfer from multiple source modellers [5]. Improved predictive performance was observed from increases in the number of sources. This report reviews the multi-source transfer findings in [5] and applies it to a real world problem of pollution modelling in India, using public-domain data.
|
|
|
Distributed Sequential Zero-Inflated Poisson Regression
Žemlička, R. ; Dedecius, Kamil
The zero-inflated Poisson regression model is a generalized linear model (GLM) for non-negative count variables with an excessive number of zeros. This letter proposes its low-cost distributed sequential inference from streaming data in networks with information diffusion. The model is viewed as a probabilistic mixture of a Poisson and a zero-located Dirac component, whose probabilities are estimated using a quasi-Bayesian procedure. The regression coefficients are inferred by means of a weighted Bayesian update. The network nodes share their posterior distributions using the diffusion protocol.\n
|
|
|
Unsupervised Verification of Fake News by Public Opinion
Grim, Jiří
In this paper we discuss a simple way to evaluate the messages in social networks automatically, without any special content analysis or external intervention. We presume, that a large number of social network participants is capable of a relatively reliable evaluation of materials presented in the network. Considering a simple binary evaluation scheme (like/dislike), we propose a transparent algorithm with the aim to increase the voting power of reliable network members by means of weights. The algorithm supports the votes which correlate with the more reliable weighted majority and, in turn, the modified weights improve the quality of the weighted majority voting. In this sense the weighting is controlled only by a general coincidence of voting members while the specific content of messages is unimportant. The iterative optimization procedure is unsupervised and does not require any external intervention with only one exception, as discussed in Sec. 5.2 .\n\nIn simulation experiments the algorithm nearly exactly identifies the reliable members by means of weights. Using the reinforced weights we can compute for a new message the weighted sum of votes as a quantitative measure of its positive or negative nature. In this way any fake news can be recognized as negative and indicated as controversial. The accuracy of the resulting weighted decision making was essentially higher than a simple majority voting and has been considerably robust with respect to possible external manipulations.\n\nThe main motivation of the proposed algorithm is its application in a large social network. The content of evaluated messages is unimportant, only the related decision making of participants is registered and compared with the weighted vote with the aim to identify the most reliable voters. A large number of participants and communicated messages should enable to design a reliable and robust weighted voting scheme. Ideally the resulting weighted vote should provide a generally acceptable emotional feedback for network participants and could be used to indicate positive or controversial news in a suitably chosen quantitative way. The optimization algorithm has to be simple, transparent and intuitive to make the weighted vote well acceptable as a general evaluation tool.\n
|
|
|
DEnFi: Deep Ensemble Filter for Active Learning
Ulrych, Lukáš ; Šmídl, Václav
Deep Ensembles proved to be a one of the most accurate representation of uncertainty for deep neural networks. Their accuracy is beneficial in the task of active learning where unknown samples are selected for labeling based on the uncertainty of their prediction. Underestimation of the predictive uncertainty leads to poor exploration of the method. The main issue of deep ensembles is their computational cost since multiple complex networks have to be computed in parallel. In this paper, we propose to address this issue by taking advantage of the recursive nature of active learning. Specifically, we propose several methods how to generate initial values of an ensemble based of the previous ensemble. We provide comparison of the proposed strategies with existing methods on benchmark problems from Bayesian optimization and active classification. Practical benefits of the approach is demonstrated on example of learning ID of an IoT device from structured data using deep-set based networks.
|
|
|
Potential Radioactive Hot Spots Induced by Radiation Accident Being Underway of Atypical Low Wind Meteorological Episodes
Pecha, Petr ; Tichý, Ondřej ; Pechová, E.
Hypothetical radioactivity release with potentially high variability of the source strength is examined. The interactions of the radioactive cloud with surface and atmospheric precipitation are studied and possible adverse consequences on the environment are estimated. The worst-case scenario is devised in two stages starting with a calm meteorological situation succeeded by wind. At the first stage, the discharges of radionuclides into the motionless ambient atmosphere are assumed. During several hours of this calm meteorological situation, a relatively significant level of radioactivity can be accumulated around the source. At the second stage, the calm is assumed to terminate and convective movement of the air immediately starts. The pack of accumulated radioactivity in the form of multiple Gaussian puffs is drifted by wind and pollution is disseminated over the terrain. The results demonstrate the significant transport of radioactivity even behind the protective zone of a nuclear facility (up to between 15 and 20 km). In the case of rain, the aerosols are heavily washed out and dangerous hot spots of the deposited radioactivity can surprisingly emerge even far from the original source of the pollution.
|
|
| |
guest ::
