Head of the Department:
Tatiana Valentine Guy

Deputy head of the Department:
Miroslav Kárný

Secretary:
Věra Králová

phone: +420 286 890 420
www: http://www.utia.cas.cz/AS
staff: people, Ph.D. students, alumni
List of publications, courses, projects

The Department of Adaptive Systems focuses predominantly on the design of decision-making systems, which modify their behavior according to the changing properties of their environment. This essential ability – adaptivity – enhances their efficiency. Decades of research have brought a number of conceptual, theoretical, algorithmic, software and application results. The applicability of adaptive systems is currently being extended toward complex scenarios by improving the classical adaptive systems and by developing their new versions.

The departmental “know-how” serves to resolve national as well as international research projects, running in collaboration with industry and government agencies. The interplay between theory and limited computing power is the common issue linking the various project domains. They include traffic control, management and control of technological systems, radiation protection, nuclear medicine, analysis of financial data, electronic democracy, etc. The increasing complexity of the problems addressed directs the main stream of the research toward decentralized control of large-scale systems and normative decision-making with multiple participants.

Last events:

Radek Hofman was awarded by the Czech Radiation Protection Society's Award

Czech Radiation Protection Society awarded prize for the best work of young authors in the field of protection against ionizing radiation to Ing. Radek Hofman, Ph.D. from AS department of the Institute of Information Theory and Automation for his PhD thesis.

Uid: 34

Cena za 1. místo v soutěži disertačních prací v jaderných oborech

Ústav jaderného výzkumu Řež a.s. a Česká nukleární společnost ocenily disertační práci s názvem "Application of Advanced Data Assimilation Methods in Off-site Consequence Assessment" Ing. Radka Hofmana, Ph.D. cenou za 1. místo v soutěži disertačních prací v jaderných oborech.

Uid: 34

NIPS Workshop Decision Making with Multiple Imperfect Decision Makers

more info: www.utia.cz/NIPSHome

You are invited to participate in the NIPS Workshop DECISION MAKING WITH MULTIPLE IMPERFECT DECISION MAKERS in conjunction with the 25^th Annual Conference on Neural Information Processing Systems, which is the premier scientific meeting on neural computation.

Uid: 49

• 1
• 2
• 3
• next ›
• last »

Department characteristics

Adaptive systems (AS) are systems making decisions or selecting control actions and concurrently improving themselves. They work under incomplete knowledge in uncertain, stochastic and dynamically changing environment. Traditionally, AS comprise adaptive estimators, detectors, predictors, controllers, etc. Design and application of AS represent long-term challenge that can be addressed only when using variety of disciplines labeled as cybernetics.

The list of people in our department indicates that the group is well balanced, covering the art of adaptive systems from theoretical, algorithmic and software aspects up to real-life applications. The group has been dealing with adaptive (control) systems and related problems more than 40 years. Through these years it has created a unified, theoretically and algorithmically well grounded approach to solving problems met in the area. The approach which can be labeled as Bayesian dynamic decision making is now perceived as Prague school of adaptive systems.

Theory and algorithms

The distinguished features of the department are:

• the research activities aim at creating a unified theory;
• the decision-making problem is solved as the technical one in its maximal possible completeness;
• the constructive approach dominates the work (the best possible solution is searched for: improvements are rarely started from the analytical side);

Applications

Applications the Department is dealing with are a source of vital feedback that directs us to real, not just 'academical' questions. They ranges from adaptive control of technological processing up to advising to human beings managing complex process in industry, economy and medicine. The energy spent on gradual building of generic algorithmic and software tools starts to pay back so that we are able to enter new application domains very efficiently.

Adaptive systems are dynamic units that learn their environment while make their decisions. Within this broad framework, the main research areas of the department are:

Probabilistic Design: Fully Probabilistic Design of Dynamic Decision Strategies

Advanced Control: Adaptive LQ Controllers and Predictive Controllers

Stochastic Sampling: Sequential Sampling Methods for Identification and Control

Linear Systems: Advanced Theory of Linear Systems

Traffic Control: Urban Traffic Feedback Control

Bayesian Decision Making: Models with strictly bounded noise

Investigated Research Topics:

Decision-Making: Adaptive Decision-Making under Informationaly Demanding Conditions

Advising: Optimized Bayesian Dynamic Advising

Multi-Participant DM: Theory of Multi-Participant Bayesian Decision Making

Advisory System: Optimizing Support Tool for Decision Makers

Urban Traffic Control: Feedback Control of Traffic in Cities

Nuclear Safety: Modeling of Consequences of Nuclear Accidents

Medicine: Support of Diagnostics and Treatment

Bayesian soft sensor: a tool for on-line estimation of the key process variable in cold rolling mills

Development of software tools was never primary aim of our research, however development of methodologies and algorithms is impossible without proper software support. At present, we are dealing with increasingly more complex systems, hence requirement on reliability and flexiblity of software tools are growing. The following projects are actively developped and maintained:

Mixtools: Toolbox for Decision Making with Mixtures

BDM Library: C++ library for Bayesian Decision Making

GPC Toolbox

Image Sequence Analysis Toolbox

Developed Software:

Jobcontrol: User Interface for Routine Use of Mixtools

Designer: Computer-Aided Design of Adaptive Controllers

LQ Toolbox

Basic Publications

The AS Department ensures, organizes and produces amount of lectures, educational materials, seminars, conferences and workshops within the domain of decision making, advanced control and related areas. The department produces a significant amount of educational material on Bayesian Decision Making. This page summarizes the information about the main educational activities held in the department.

Educational materials

Topics for students

Seminars

Workshops

University courses

Mailing address:
Department of Adaptive Systems
Institute of Information Theory and Automation 
P.O. Box 18
182 08 Prague 8
Czech Republic 

Visiting address:
Pod Vodárenskou věží 4
Prague 8 - Libeň
Czech Republic

Tel: +420 286890420
E-mail: school@utia.cas.cz

Detailed info how to reach the building can be found here.  

Knowledge extraction maps extensive data sets on lower dimensional objects. Its results always serve to a subsequent, often dynamic, decision making. Decision-making quality is substantially influenced by the mapping used. This simple fact is relatively rarely respected by many elements in the overwhelming arsenal of existing mappings. A complete solution of decision making problems that includes explicitly the discussed mapping are severely limited by computational complexity (labelled as curse of dimensionality).  The project contributes to an improvement of this state via

i)   solving general dynamic-decision tasks within a specific Bayesian methodology that uses probabilistic tools  both for describing the object and strategies of decision making but also its aims and constraints;

ii)  developing methodology approximating the optimal solution obtained;

iii) verifying the developed methodological and algorithmic tools on non-trivial, practically significant, decision-making problems in medicine (diagnostics of secondary lymphedema) and economy (trading with futures).

Contact:

M. Kárný

Dynamic decision making (DM) maps knowledge into DM strategy, which ensures reaching DM aims under given constraints. Under general conditions, Bayesian DM, minimizing expected loss over admissible strategies, has to be used. Existing limitations of the paradigm impede its applicability to complex DM as:

1. Complexity of the information processing often crosses resources accessible.
2. Quantification of domain-specific knowledge, aims and constraints is weakly supported. It concerns mapping of domain-specific elements on probabilistic distributions (pd).
3. Methodology of the DM with multiple aims is incomplete.

The research aims to overcome these problems. It relies on distributed DM and fully probabilistic design (FPD) of strategies. The goal is to build a firm theoretical background of FPD of distributed DM strategies. Besides, it will enrich available results and unify them into internally consistent theory suitable for a flat cooperation structure.

This aim implies the main tasks:

1. Inspection of conditions leading to FPD
2. Extension of FPD to design with sets of ideal pds
3. Design of computerized conversion of knowledge and aims into environment-describing and ideal pds
4. Elaboration of theoretical framework for selecting cooperation tools

Contact:

M. Kárný

This research direction is concerned with identification and control of uncertain systems using Bayesian decision-making theory. The main advantage of this theory is consistency of the generated decision (i.e. estimates and control actions). However, solution of the implied recursive Bayesian relations is often available only approximately. 

Sampling methods provide a traditional approximation methodology for Bayesian statistics. Any complex probability density function can be approximated by a set of samples generated from it. This method is computationally expensive, however research effort to increase efficiency of sampling methods and increasing performance of computers improved applicability of these methods in such a way that they bring significant improvement in many application areas and represent a strong alternative to traditional approximation methods.

Sequential Monte Carlo is a way to apply sampling methods for on-line estimation and filtering. It is an established methodology with many practical applications. The advanatage of the methodology is it universality with a possibility to tailor the algorithm for a particular problem via proposal density or Rao-Blackwellization. 

A new research direction is based on application of particle filtering methods in control. This is based on duality between estimation and control. Particle filters can be applied to both dynamic programming and model predictive control formalization of the control task.

Various specific features of the approch are being elaborated under nationally funded projects listed below.

Contact:

Projects:

Despite of huge progress in intelligent control, neural networks, fuzzy and nonlinear control, and other parts of control theory, the methods of linear control still remain a basis the other theories are compared with. Linear models have proved to be a relatively easy but powerful tool that has been successfully applied to many problems at work, and which has reached a high level of development during the last decades.

The lectures on linear systems and control also form a core of university courses devoted to systems theory and control. Nevertheless, the existing open problems show that there is still room for further growth and improvement of existing methods and inventing new approaches and methods.

The main goal of these studies is to contribute to the development of new methods and algorithms for the analysis and synthesis of linear control systems (with constant parameters and with or without time delays). An important vehicle for meeting the goal is the exploitation of numerous theoretical works of the recent period, for example our own contributions pertaining to the problems of matrix completions of polynomial matrices.

Contact:

The research deals with urban traffic feedback control systems. The general objective of the project is the enrichment of the complete design line of LQG controllers so that it will cover steps related to state estimation, ideally with mixed-type (continuos and discrete) states. The most important problems expected to be solved during the project are (i) general solution of state estimation in factorized form and its specialization to linear Gaussian state-space models (ii) translation of the users knowledge into optional parameters of the resulting factorized filter (iii) design of filters estimating mixed-type states and (iv) implementation and testing of controllers with state estimation on realistic simulation of traffic control problem.

Contact:

E. Suzdaleva

The main direction of the research is a design and investigation of model-based control approaches and methodologies for their real implementation and self-tuning of their parameters.

• Adaptive Linear Quadratic Control
• Generalized Predictive Control

The issue of the model-based approaches is

1. A choice and composition of suitable model for given control strategy.
   The models taking into consideration are:
   • Input Output regressive models
   • State-Space models
   The models are obtained either by some identification procedure (least squares, probabilistic theory) or they are determined on the basis of mathematical-physical analysis.
2. Tuning of the controllers.
   The tuning is investigated from two sides:
   • aimed advices for manual tuning based on mathematical-physical relations
   • automated tuning based on probabilistic theory

History of AS department

AS department was created in middle of sixties of the past century. Control applications based on physical modeling reached soon barrier that stems from complexity of the constructed models and impossibility to find feasible controllers to them. It was found that simple black-box models are often sufficient for design of efficient controllers. The need to learn model structure and its parameters stimulated interest in so called experimental identification. Search for an adequate methodology gradually singled out Bayesian methodology as the only known systematic tools suitable for solving the addressed class of problems. Gradually, following the improvements of the theoretical, algorithmic and evaluation tools, the interests have shifted to multivariate, non-linear and non-Gaussian cases. Also, control of basic level of technological processes has been gradually substituted by  higher level control and other application domains (physics, medicine, economy, societal decision making etc.). Attempt to created applicable generic tools and struggle with curse of dimensionality has become the main driving forces of the research we perform.

During decades of research a lot of people and partners contributed to our current know how, see the alumni list and list of honorary members. It is also worthwhile  to scan workshops and seminars we organized: they clearly demonstrate both paradigm shift we underwent including circles we return back to old ideas and old problems.  The  list of people actively working within the department, the recent seminars and addressed research as well as application topics indicate that the department is flourishing and contributes to progress of the field.

Alumni

Previous grants and projects

The algorithmic and software implementation of theory of optimized Bayesian dynamic advising served as a basis for construction of advisory system intended to support the decision-maker.

To customise a particular advisory system, a large sample of historical data taken from managed process is analysed and processed offline. The obtained results are complemented by information about the expected advisory levels and decision-making aims.

A core of the advisory system forms Mixtools package, which has been implemented both: as a toolbox within MATLAB environment and as MATLAB-independent code. The MATLAB-like implementation is intended to serve to research and simulation purposes. Another implementation can be integrated with an existing control and/or monitoring system of the process managed and, thus, can serve to real-time, full-scale application.

The advisory system was implemented and extensively tested on several different case studies: prediction of urban traffic, treatment of thyroid gland carcinoma and fault detection and isolation problem. A real-time, full-scale industrial implementation of the advisory system on cold rolling mills confirmed the generic nature of the tool and illustrated the following key features of the solution:

• ability to tailor to specific system managed
• ability to support permanent adaptation
• ability to follow the best practice available
• ability to provide the strategy leading to the desired behaviour of system managed
• ability to cope with multiple decision-making aims
• ability to control the information load on the decision-maker (the techniques forming the core of advisory system are hidden from the decision-maker and advices are presented in easy-to-understand graphical form)

The system and its core Mixtools package are permanently innovated and improved. For the latest version, please, contact M.Kárný.

Contact:

M.Kárný 

Support of grants

• Dynamic clustering: theory, algorithms and software, Grant Agency of the Czech Republic CR No. 102/03/0049, 2002-2005
• Dynamic clustering for control of complex processes, Academy of Sciences of the Czech Republic No. S1075351, 2002-2005

Local SVN repository (accessible with password only):

http://mys.utia.cas.cz:1800/svn/mixtools

The task aims at building a multi-level control of the traffic in large urban transportation nets. The basic unit we operate with is the traffic microregion. It is a logically delimited collection of crossroads and the communications joining the crossroads. We suppose, some of the crossroads are controlled by signal lights and the arms of the controlled crossroads are equipped by detectors -- measuring devices, providing us with transportation data (intensities and densities of the traffic flow).

The basic variable we model and control is a vector of queue lengths that are being formed in the arms of the controlled intersection. The queues are approximated by the number of vehicles, using the physical principle "the increment of the queue is given by the difference in the amount of incoming and outgoing cars". In addition to this, a linear dependence of the car density measured on the remote detector on the column length is considered. In this way a state space model for queue lengths and density in the microregion is constructed.

The controller built on the basic of the presented model has three levels:

1. Local control which cares of a single microregion, only. It predicts column lengths of the microregion (which are no measurable) and computes ratios of the greens for signal lights, so that the columns would be minimal. The estimation is performed by Kalman filtering, for the optimization is used linear programming.
2. Coordinating control cares about the "good relations" between the microregion. With no coordination, the microregions incline to overload their neighbors, if convenient for them. The coordination prevents this phenomenon. Both the model and the optimization are just slight modifications of those from the local control.
3. Exceptional states control cares about the states of the traffic, which do not belong to the common transportation. They are e.g. accidents, closures, big over-saturations etc. They are supposed to be described by discrete model and through them, mostly control using the collected expert knowledge will be realized.

The traffic control algorithm based on these principles has been developed in cooperation with ELTODO EG and finalised and practically tested in cooperation with ELTODO dopravní systémy.

Currently we are in the process of extending the model and improving the controller in the scope of a project NOMŘÍZ, a joint effort of ÚTIA, ELTODO dopravní systémy and Czech Technical University, Faculty of Transportation Sciences.

Contact:

• Jan Přikryl
• Jitka Homolová
• Ivan Nagy

Projects

• Ministry of Transportation of the Czech Republic, No. 1F43A/003/120,
• Ministry of Education, Youth and Sports of the Czech Republic, No. 1M0572 (Centre of applied research DAR ),
• Technology Agency of the Czech Republic, No. TA01030603 (NOMŘÍZ).

Local SVN repository (accessible with password only):

http://mys.utia.cas.cz:1800/svn/doprava

Assessment of radiological impact of accidental and normal radioactive releases on population. Application of multi-pathway transport model for regulation of normal atmospheric radioactive discharges from nuclear facilities. Advance from deterministic assessment of radiological consequences of radioactive releases into atmosphere toward the probabilistic approach.

Adaptation of techniques enables progress from former deterministic calculations towards the generation of probabilistic answers on assessment questions. Uncertainties of input parameters are taking into account and their propagation through the mathematical model is treated. Adopted scheme of Monte Carlo modeling uses stratified sampling procedure LHS. Uncertainty analysis and sensitivity analysis techniques are used to classify the extent of the uncertainty on predicted consequences and rank of particular input parameters according to their influence on radiological endpoint values.

Development of the proper sequential data assimilation techniques for corrections of model predictions on basis of observed (measured) values in terrain. Verification of various minimization algorithms with regard to complex task of radionuclide propagation into the living environment. Cooperation on development of interactive user friendly program tool customized for conditions of nuclear facilities in the Czech Republic for support of decision making during nuclear emergencies.

Contact:

In this domain we are oriented towards computerized support of difficult diagnostic and treatment problems. The particular important applications serve us as test-bed of our generic methodology, algorithms and software tools. The specific problem of nuclear medicine topics is a lack of data for processing which points out the advantage of prior information.

Nuclear medicine - thyroid cancer

• Modelling of activity-time course in thyroid gland after administration of ¹³¹I to a patient,
• Advisory system for patient-individual recommendation of therapeutic activity of ¹³¹I.

Nuclear medicine - lymphoscintigraphy

A system for early diagnosis and treatment of upper limb lymphedema using quantitative lymphoscintigraphy

Internal medicine

Medical Imaging methods used for diagnostics in internal medicine.

Model of activity kinetics

After oral administration of radioactive ¹³¹I to a patient, iodine is accummulated in thyroid gland. Its activity rapidly increases and then slowly decreases. Model of activity time course is useful for (i) prediction of thyroid activity in a near future, (ii) time integral of activity is proportional to a dose (i.e. energy of the radiation) absorbed in the tissue. Thyroid activity is measured once or twice a day and usually not much more than 3 such measurements are available. Furthermore, these data contain random and potentially other measurement errors. Because small amount of uncertain data, prior information has been balanced to decrease uncertainty of estimated model parameters and, on the other hand, not to overweight information carried by the data.

The time integral of thyroid activity has been estimated as a random quantity. Its distribution is used for dosimetric and radio-hygienic purposes and as an input quantity for statistical analyses as well.

Advisory system

Radiodestruction of thyroid tumour is achieved by administration of ¹³¹I with high activity. The aim is to destroy the target tumour but, on the other hand, to minimize secondary radiation risks. As the response of patients' organism to administered activity is individual, therapeutic activity must be administered individually as well.

The advisory system is based on probabilistic mixture describing a selected multidimensional subset of characteristic patients' data. Then the advisory mixture model is designed, reflecting the user request to minimize the administered activity with successful result of therapy. Advice for a specific patient is conditioned by his actual data obtained in diagnostic examination before the planned therapy.

The results demonstrate that it is crucial to collect wide enough data set for description by the probabilistic mixture. The advices corresponded to medical decisions in one category of the disease that was sufficiently described by the available data used for mixture estimation.

Next effort is focused on

• collection of more data (variables and records),
• inclusion of incomplete data records into mixture estimation,
• dealing with mixtures of both continuous and discrete variables.

Contact:

L. Jirsa

Support of grants

• Intelligent decision support of diagnosis and therapy in nuclear medicine by Bayesian processing of uncertain data and probabilistic mixtures, Academy of Sciences of the Czech Republic, 1ET1 007 50404, 2004-2007
• Centre of applied research DAR, Ministry of Education, Youth and Sports of the Czech Republic, 1M6 798 555 601
• Dynamic clustering: theory, algorithms and software, Grant Agency of the Czech Republic, 102/03/0049, 2002-2005

Quantitative lymphoscintigraphy for diagnostics and therapy of upper limb lymphedema

Inspection by lymphoscintigraphy is potentially the method searched for. Its potential for examination of upper limbs is, however, inhibited by the lack of a reliable quantitative evaluation when upper limbs are examined. The main reason is number of measurements limited both by the time-capacity of the gamma camera and by the ability of a patient to undergo a series of measurement in long time intervals.

The general aim of this project is to develop new automated lymphedema diagnostics based on combination of scintigraphy quantification and other indicators available. This task can be divided into two tasks:

1. To develop routinely applicable quantitative evaluation of the state of upper limbs. The Bayesian paradigm, in conjunction with a simplified model of the diffusion dynamics, is used to obtain reliable quantitative evaluations for the first time. Among the resulting benefits are:
   1. a procedure for tuning the measurement conditions,
   2. an assessment of deviations between limbs,
   3. the possibility of lymphedema staging assessment.
2. To combine developed scintigraphy quantification with the rest clinical and qualitative scintigraphy findings into one global automated lymphedema stage classifier in order to increase the accuracy of diagnostics. The advantages of such global classifier are:
   1. a maximum information available for the decision,
   2. a minimized impact of skill of inspecting expert on the disease assessment,
   3. the possibility to evaluate the relevance of subset of features on the disease assessment; The relevance of the developed quantification in the combined diagnostics can be verified in this way.

Diagnostic value of the developed combined lymphedema diagnostics will be tested on the real data within the full diagnostic-therapeutic cycle and compared with conclusions of alternative diagnostic evaluations.

Contact:

Application in Internal Medicine - Computer Image Diagnosing

Methods of medical image diagnosis are developed in AV CR project 1ET101050403. These methods are based on modelling of healthy and unhealthy tissue image pattern features using Gaussian mixture models. Decision making is based on Bayesian framework.

Internal medicine diseases are diagnosed.

• Successful application in Endocrinology: diagnosing thyroid gland disease (Hashimoto's Lymphocytic Thyroiditis) with ultrasonography
• Planned application: diagnosing liver tumor with computer tomography

Contact

M.Kárný

Support of grants

• AV CR 1ET101050403
• Centre of applied research DAR, Ministry of Education, Youth and Sports of the Czech Republic, 1M6 798 555 601

Mixtools is a toolbox designed for learning, prediction and control design with probability mixtures with a stress on fully probabilistic of strategies. The toolbox functions cover:

• Initialization of mixture estimation.
• Approximate mixture parameter estimation.
• Prediction with mixtures.
• Design of adaptive decision making strategies
• Design of advisory systems
• Visualization.

The toolbox functions offers the possibility of processing with high-dimensional data records, dynamic mixture components and extensive sets of learning data.

Contact:

Support of grants

• Fully probabilistic design of dynamic decision strategies, Grant Agency of the Czech Republic, No. 102/08/0567, 2008-2011
• Research center DAR, MŠMT 1M0572, 2005-2011
• Dynamic clustering: theory, algorithms and software, Grant Agency of the Czech Republic No. 102/03/0049, 2002-2005
• Dynamic clustering for control of complex processes, Academy of Sciences of the Czech Republic No. S1075351, 2002-2005

SVN repository (with password only):

Jobcontrol is a user friendly interface for Mixtools and Designer toolboxes. The Mixtools toolbox is a powerful set of utilities for system identification employing mixture models and the corresponding control design. It is implemented as set of M-scripts and MEX-binary exacutables for the Matlab computing environment. It suits to the goal of finding suitable structure for given data. The Designer toolbox then serves for finding optimal controller parameters, constructing ideal controller and testing the controller found.

As an expert tools, Mixtools and Designer fullfils end user's needs, but are not totally suited for direct usage of the end user. In other words, they are not very user-friendly. It is why, we are developing, environment, which integrates all the tasks, that are connected with system identification and controller design and helps to collec all the user's knowledge of data and the real-world system where data come from. The Jobcontrol package, therefore, integrates endless expertise that is otherwise available only through study of the theoretical books - Optimized Bayesian Dynamic Advising: Theory and Algorithms by M. Karny et al. and [P. Nedoma, M. Karny, T.V. Guy, I. Nagy, and L. Tesar. Learning and prediction with normal mixtures. Technical Report 2045, UTIA AV CR, 2002], Mixtools toolbox documentation Bayesian approach to system identification by V. Peterka and experience contained in many experiments.

The Jobcontrol package help to solve user's problem in terms of the experiment (or job). Every experiment consists of description of user's data, and description of the way how the mixture is estimated and how the control is performed and what tests are to be done. Jobcontrol offers the user environment for interactive input of the description of experiment as well as lucid way of configuring experiment using one cnfiguration file. Integral part of Jobcontrol package is the protocol generator, which automatically creates a very convenient LaTeX document, which shows all the aspects of system identification, control and user's data description.

Contact:

SVN repository (password only):

Controller tuning is a basic step in any control application. This tuning is a complex process composed of several steps starting with the plant analysis and ending with the verification of the designed controller. There exist various tools that help in particular steps of the design but the complete path of the design is not supported. This work makes an attempt to offer a procedure of "complete" controller design where all necessary steps follow automatically one after another. The idea is applied here to the LQG controller design. The whole procedure is described and demonstrated on an example with the emphasis on the tuning of the LQG criterion to respect the given constraints.

The steps of the Designer are:

1. Data preprocessing
2. Structure estimation
3. Parameter estimation
4. Closed-loop optimization
5. Verification

Currently, the Designer toolbox is merged with the Mixtools where it can be accessed for example using the Jobcontrol interface.

Contact:

SVN repository (password only):

The library is designed using object oriented approach where decision-making is implemneted as a method of dedicated object: decision-maker. The library contains many commonly known decision-makers such as estimators and Bayesian filters. Support for control-oriented decision-makers (LQG control) is under development.

Design philosophy of the toolbox is tocreate a close image of the underlying theory. The library is build from objects representing random variables, probability density functions (pdfs) and Bayesian models. Calculus with probability density functions is implemented eiter as:

• methods of classes representing pdfs: marginalization and conditioning
• special types of pdf classes: the chain rule
• and operation of class Bayesian model: the Bayes rule.

The library also contain supporting classes for running experiments with Bayesian decision makers, such as:

• loggers: classes for recording results and intermediate results of the experiement in various formats.
• user info: classess for loading and storing experiment setup (parameters, conditions, etc.) in XML for easy editation by the user.

For more information see project page: http://mys.utia.cas.cz:1800/trac/bdm

Contact:

Support of grants

• Centre of applied research DAR, Ministry of Education, Youth and Sports of the Czech Republic, 1M6 798 555 601
•  GA ČR 102/08/P250, 2008-2010

LQ toolbox was created out of the need to demonstrate the characteristics and application of

• LQ controllers, which require knowledge of the discrete tranfer function of the controlled system
• Adaptive LQ controllers.

• Simulink schemes for
  • a fixed model controller ( various schemes to demonstrate various characteristics)
  • an adaptive version of LQ controller for several types of models : standard regression model, incremental model (with integrator factor), both versions with or without measurable external disturbance and the possibility to consider pre-programming ( use future values of reference signal)
• Simulink blocks representing above mentioned types of LQ adptive controller
• m-files for LQ optimization and identification

Author:

Contact (info):

FTP:

GPC toolbox serves for obtaining the basic knowledge about the Generalized Predictive Control (GPC). It is prepared for control experiments of Linear Single-Input Single-Output systems with Time-Invariant parameters (LTI SISO systems) described by Input Output differential equation or state-space form.

The GPC toolbox enables user to study the properties of the basic algorithm, generating full control actions and incremental predictive algorithm. The toolbox is prepared in two identical versions:

• version based only on MATLAB scripts
• version combining Simulink schemes and MATLAB scripts

MATLAB scripts and Simulink schemes make possible to select and to change

• input parameters of the controller
• trajectory (desired reference)
• controlled system
• magnitude of immeasurable disturbance of the output

The both versions offer a lot of different possibilities of diagnostics of the control process.

Contact:

Link:

The book provides the theoretical and algorithmic basis of optimized probabilistic advising. Starting from abstract ideas and formulations, and culminating in detailed algorithms, the book comprises a unified treatment of an important problem of the design of advisory systems supporting supervisors of complex processes. It introduces the theoretical and algorithmic basis of developed advising, relying on novel and powerful combination black-box modeling by dynamic mixture models and fully probabilistic dynamic optimization.

Decision Making with Imperfect Decision Makers by T.V. Guy, M. Kárný, D.H. Wolpert, Springer, 2012

The educational materials on Bayesian decision making produced in AS Department are presented at this page. The material is organized so that to be potentially useful for different target groups of users: from students and PhD students to engineers solving practical problems. Textbooks are listed in the basic publications of the AS Department.

Dynamic Decision Making: Fully Probabilistic Design

The presented lecture in form of slides is the most up to date material. It provides a unified basis of dynamic decision making under uncertainty and incomplete knowledge. A package of examples to this lecture will be developed soon.

Bayesian Decision Making: Theory and Examples

This part of educational materials provides Bayesian decision making theory for beginners. It includes basic theoretical materials and examples available to download.

Examples on Bayesian Decision Making

This part of educational materials offers Bayesian decision making for experienced researchers and engineers. The provided examples deal with the research carried out in AS Department. Most of them are implemented in toolbox Mixtools.

The introductory part to Bayesian Decision Making deals with four basic tasks:

• modelling (simulation),
• estimation,
• prediction
• and control.

These tasks, used for single input - single output cases, are simple enough to demonstrate clearly the basis of the whole theory and, on the other hande, they are mostly needed and used in the practice. The basic theory results to algorithms which are implemented in Octave (open source clone of MATLAB). The examples are available to download in the svn repository.

Contact:

Support of grants

• Educalibre 110330-CP-2003-ES-Minerva-M

SVN repository (with password only):

This part of educational materials offers Bayesian decision making for experienced researchers and engineers. The provided examples deal with the research carried out in AS Department. Most of them are implemented in toolbox Mixtools available for download in the svn repository.

Contact:

SVN repository (with password only):

Aims of Edukalibre

Partners

AS participation

Contact:

Support of grants

The project has been supported by the following grant:

• European Commission, No. 110330-CP-2003-ES-Minerva-M

Local SVN repository (accessible with password only):

Our department is actively involved in activities of Czech Society for Cybernetics and Informatics (CSKI). Specifically, one its group "Decision-Making and Control under Uncertainty" (DCU) was founded by members of AS department. The aims of this group cover the main research interests of our department.

Moreover, local seminars are organized within regular meeting of members of the department every monday. Primary role of the local seminars is communication of knowledge within the department, everyone is welcome to attend these seminars. Information about "Monday seminars" are dissipated via email list seminar<at>utia.cas.cz.

To subscribe to this list, send a message to listproc@utia.cas.cz with content:

"SUBSCRIBE SEMINAR <Your Name>"

List of past seminars: AS and CSKI

 Workshop:

Previous workshops:

• NIPS 2011 workshop on Decision Making with Multiple Imperfect Decision Makers

• NIPS 2010 workshop on Decision Making with Multiple Imperfect Decision Makers

• PhD workshop 2009, Young Generation Viewpoint

• PhD workshop 2006, Young Generation Viewpoint

• CMP'04: MULTIPLE PARTICIPANT DECISION MAKING

• Information Technologies and Control. Young Generation Viewpoint 2003

• Cybernetic & Informatics Eurodays. Young Generation Viewpoint 2000

• The 3rd IEEE European Workshop on Computer-Intensive Methods 1998

• IEEE European Workshop CMP'96

• Summer School Prague'95

• Dycomans Workshop Prague'95 (Organized in cooperation with Control Theory Department)

Linear Quadratic Control (LQ Control) investigated in the department consists in minimization of quadratic criterion by dynamic programming. The adaptive character of the control is achieved by addition of on-line identification of controlled system. In the initializing step of the identification, the structure of identified model is determined and the first setting of model parameters is done. During the run, the identification improves individual model parameters.

Computer-aided design and self-tuning

The challenge of the research is a design of self-tuning for the parameters of Linear Quadratic Gaussian Controllers (LQG Controllers). The tuning is a complex process composed of several steps starting with the plant analysis and ending with the verification of the designed controller.

Author:

Contact (info):

Software tools:

Generalized Predictive Control (GPC) is a multi-step approach. It combines feed-forward part and feedback part. The feed-forward part is represented by prediction via mathematical model describing a controlled system. This part forms the dominant part of control actions. The feedback, closed from measured outputs, compensates some inaccuracies of the model and certain bounded disturbances.

The real design consists in composition of equations of predictions and minimization of quadratic criterion, in which the equations of predictions are involved. The minimization is performed within finite horizons.

The research is focused on state-space control design applied to deterministic linear systems, deterministic nonlinear systems and slightly stochastic systems. Developed control algorithms are tested on mechanical systems as industrial robotic structures.

Basic algorithms of predictive control are available in GPC toolbox for MATLAB&Simulink. The toolbox contains both m-functions and c-coded functions and Simulink schemes.

Authors:

Contact:

Software tools:

Textbook:

About this textbook:

Quality of maintaining of complex man-machine systems very much depends on experience, skills and performance of the human decision-makers (operators) managing the system. The task is complicated by complexity and dimensionality of the system managed as well as limited abilities of the operator.

The research concerns developing prescriptive theory of Bayesian dynamic decision-making (DM) under uncertainty that allows to construct efficient adaptive DM systems and to create systems supporting human decision-makers. The adopted approach relies on black-box modeling and on the availability of informative data. Specialization of the developed theory to dynamic mixtures combined with fully probabilistic design provides a practical tool of broad applicability.

The general idea is to process historical data available to model of the managed system behavior under various working conditions in a form of multi-dimensional probability mixtures (learning phase). The mixture learned and mixture expressing DM aims are employed to build an advisory mixture describing DM strategy (design phase). Decision designed by an advisory system is the prediction of advisory mixture made for the actually incoming data. Advising supposes providing this prediction in a suitable form to the decision-maker. The decision-maker is responsible to accept or to reject the offered advise.

The established solution has proven to able to cope with dynamically changing incompletely known multi-attribute environment and to learn and optimize dynamic decision-making strategy realized either by human being or automatically.

The developed generic optimized dynamic advising covers:

• concise problem formulation
• general solution of learning phase, complemented by its detailed elaboration to normal and Markov-chain mixtures
• solution of learning initialization problem
• general solution of the design phase, complemented by its detailed elaboration to normal and Markov-chain mixtures

The developed theory has been practically implemented into algorithmic and software toolsets (Mixtools) and tested on several full-scale applications (see Advisory system).

Distributed dynamic decision-making and learning under uncertainty in complex and changing situations are emerging as the key competencies required to support future information-based systems. The Bayesian paradigm is acknowledged to provide a consistent and rigorous theoretical basis for joint learning and dynamic decision-making. The established theory already provides a class of efficient adaptive strategies. However, this approach fails to overcome the computational complexity barrier encountered in complex settings. This project aims to create a theoretical and algorithmic basis of a mathematically rigorous, but computationally tractable Bayesian distributed dynamic decision-making system, fully scalable in the number of local decision makers.

Objectives

The project aims to develop theory, algorithms and software for Bayesian distributed dynamic decision-making. It will make a qualitatively new step towards a generic theory of multi-participant, multi-step decision making in complex dynamic situations. The project will transform the theory into a generic algorithmic and software toolset.

The theory and its conversion into a practical tool will provide: 

1.  Rigorous learning methods applicable both to quantitative and qualitative data. 
2.  Learning of dynamic, mixed-data,probabilistic mixtures. 
3.  Computationally tractable approximation to the fully probabilistic design of decision-making strategies. 
4.  Design of a tool set for tailoring of the generic decision-making tools to a specific application. 
5.  Numerically robust algorithmic counterparts of theoretical operations. 
6.  Structured training materials for users of a varying level of competence and background. 
7.  Portable software implementation allowing accommodation of new elements developed in the heterogeneous research situation.

Applications to non-trivial problems will be used to measure the project?s success. Simulation, pilot-plants and real-life (in rolling mill industry) tests will serve this purpose.

Contact:

M. Kárný

Bayesian Dynamic Decision Making

Dynamic decision making (DM) maps knowledge into DM strategy, which ensures reaching DM aims. Under general conditions, Bayesian DM, minimizing expected loss over admissible strategies, has to be used. Long-term research covers: i) theoretical support whole design leading to fully probabilistic design generalising Bayesian DM; ii) support of particular steps of DM, like structure estimation; iii) formulation and solutions specific tasks as like probabilistic support of operators or trading with futures; iv) algorithmisation facing, for instance, poorly informative data, local nature of models, approximate estimation of dynamic mixtures; v)  distributed DM, especially, performed by decision makers with limited cognitive abilities.      

Advanced Control

The inspected problems ranges from extansions of  linear control theories, over adaptive, numerically robust, linear-quadratic control, its extension to predictive controllers oriented towards mechatronic systems. The progress is driven by advanced applications oriented, for instance, towards rolling mills or electrical motors. They call for ellaborating various technigues like: i) non-linear filtering based on marginalised particle filtering or design of soft sensors; ii)  control design for specific, say, mechatronic systems or universal controlled high-dimensional dynamic mixture models; iii) inspecting dual and distributed variants of control desing.

Traffic Control

Strong research group oriented on traffic-control domain covers theoretical, algorithmic and application-specific aspects like: i) traffic-lights based adaptive hierarchical control of town traffic; ii) estimation of an exact position of vehicle facing GPS inaccessability; iii) personal advanced system supporting economical driving style.  

Nuclear Safety

Advanced physical modelling, custemisation of general techniques to Czech teritory, tailoring of advanced Bayesian technique for data asimilation, specific algorithms fighting with problem dimensionality are key techniques developed and used for solving nuclear safety problems, especially, for advising to authorities in case of nuclear (possibly chemical or even terroristic) threats.      

Medical applications

Bayesian techniques have been traditionally developed, refined and applied in variaty of medical, predominantly dignostics, problems.  Beign improtant on their own, they have served as test field with a a alck of universal (physical-like) models, very limited amount of very uncertain measured data of mixed nature and significant consequences for patients with dangerous diseases like thyroid gland cancer or lymphedema. 

Accumulation of Acquired Knowledge 

Complexity of the targeted research makes care about knowledge accumulation in software and educational material its indispenseable part.  

List of former application-oriented projects

Dynamic decision making (DM) maps knowledge into DM strategy, which ensures reaching DM aims under given constraints. Under general conditions, Bayesian DM, minimizing expected loss over admissible strategies, has to be used. Existing limitations of the paradigm impede its applicability to complex DM as:

1. Complexity of the information processing often crosses resources accessible.
2. Quantification of domain-specific knowledge, aims and constraints is weakly supported. It concerns mapping of domain-specific elements on probabilistic distributions (pd).
3. Methodology of the DM with multiple aims is incomplete.

The research aims to overcome these problems. It relies on distributed DM and fully probabilistic design (FPD) of strategies. The goal is to build a firm theoretical background of FPD of distributed DM strategies. Besides, it will enrich available results and unify them into internally consistent theory suitable for a flat cooperation structure.

This aim implies the main tasks:

1. Inspection of conditions leading to FPD
2. Extension of FPD to design with sets of ideal pds
3. Design of computerized conversion of knowledge and aims into environment-describing and ideal pds
4. Elaboration of theoretical framework for selecting cooperation tools

Active projects

Former projects

	From	Till
Fully probabilistic design of dynamic decision strategies	01.01.2008	01.01.2011
Advanced techniques of Bayesian decision making in complex systems	01.01.2004	01.01.2009
Fully probabilistic design of adaptive decision-making strategies suitable under informationally demanding conditions	01.01.2006	01.01.2009
Bayesian Decision Making to Support Change Detection in Complex Manufacturing Systems	01.01.2007	01.01.2008
Stochastické modelování dynamických kooperativních sítí	01.01.2007	01.01.2008
Data-driven modelling for decision-making support and process monitoring	01.01.2005	01.01.2006

Electrical drives are part of everyday world. While the technology for their control is well known and reliable, new challenges are comming with new technology and new requirements. The always present pressure for better reliability, safety and cost of production and operation are the driving force for inovation.

The electrical drives are also good laboratory to test new theoretical results. We apply the results of reasearch in areas of:

• stochastic sampling methods,
• advanced control,

The resulting algorithms help us to improve reliability of the drive in low speed regimes.

Contact:

V.Smidl

Car transport: Fuel consumption optimization

This research project aims at optimization of fuel consumption both from the economical and ecological points of view.

Contact:

Ecology of driving

The project aims at creation an advisory system for drivers of vehicles or other transportation means. The advisory system will measure the traffic variables and in dependence on them it will evaluate the ecological quality of driving. The advisory system will be based on a cluster model describing the data space. In this space, the data vectors are represented by points. The model describes and evaluates the point cluster that represent individual working regimes. According to the evaluation of these clusters, the quality of driving is assigned.

Contact:

Estimation of Vehicle Trajectory during GPS Signal Outages Based on Inertial Sensor Data

The aim of joint project under the research centre DAR   is to refine the information about the moving vehicle position obtained from global positioning system (GPS). In the case of signal outage, GPS does not work properly. During this outage, the vehicle position is estimated. The estimation is based on the principle of the inertial navigation (INS). Data required for the estimation are obtained from the following sources: (i) controller-area network (CAN) that provides a data from the vehicle sensors; (ii) external device that provides an acceleration and an angular velocity in the three axes from MEMS sensors; (iii) GPS navigation. All above mentioned devices are tightly connected with the vehicle.

Contact:

L. Pavelkova

HMI – Human Machine Interface

Industrial case studies focused on an on-board HMI supporting interactions between driver and on-board systems, e.g. for measuring quantitative and qualitative parameters of driver behaviour. This area of research is mainly focused on investigation driver bio-physiological parameters including driver decoy. Development systems providing information about actual, vehicle-related and predicted HMI parameters based on communication on car bus with support HMI Simulator. HMI simulator is understand as equipment, on which is possible in laboratory way objectively measure driver stress induced by different disturbing effects and subsequently judge influence of stress on safety factors.

Contact:

Industrial partners:

A state space model is frequently used for a description of real systems. Usually, some state variables are hidden and cannot be measured directly and some model parameters are unknown. Then, the need for learning, i.e., the state filtering and parameter estimation, arises. Probabilistic models provide a suitable description of the always uncertain reality and call for such approaches as Bayesian learning. Uncertainties are standardly modelled by the Gaussian distribution. This leads to Kalman-filter-based algorithms.

However, the modelled quantities are often physically constrained. Then, methods based on the Gaussian distribution with unbounded support do not work properly and they have to be adapted. The alternative sophisticated algorithms based on “unknown-but-bounded errors” principle address the same problem but they are poorly harmonised
with the subsequent dynamic decision making (like control, prediction of hidden quantities or future measurements) to which any learning serves.

This research operates in probabilistic framework while coping with bounded uncer-tainties and physically constrained quantities. Here, learning algorithms for models with constraints are constructed that (i) are based on the Bayesian principle, (ii) are recursive and (iii) have relatively simple setting and maintenance, (iv) are at disposal to subsequent
dynamic decision making.

Contact:

L. Pavelková

One of the key objectives of any rolling mill control system is to keep the thickness of the processed material within the prescribed tolerance band, which can be as low as +-10 micrometers for thin strips. Failure to comply with the tolerances results in losses which, according to experts estimate, might go up to 10% of the profit for poorly equipped rolling mills. Unfortunately, no practical direct measurement of the gauge within the rolling gap is possible. Strip thickness can be measured 50--100cm after the rolling gap with a high transport delay (20--120 samples). Thickness measurement devices minimizing or eliminating the delay are very expensive, therefore other ways of output thickness estimation/prediction are utilized. The rolling process is modelled using either well-established time-proven principles or simple black-box linear regressive models. The models are treated as probablilistic mixture and their thickness predictions are merged by the dynamic model weights in the mixture, respecting instantaneous operating state of the plant. At the same time, measurement methodology of several quantities (speed, pressure, ...) has been revised and improvement in precision and reliability indication has been achieved.

This application is a joint effort of ÚTIA, Compureg Plzeň, Jožef Stefan Institute and INEA, both from Ljubljana, Slovenia.

Kontakt: Jirsa, Dedecius

V. Peterka, J. Krýže, and A. Fořtová. Numerical solution of Wiener-Hopf equation in statistical identification of linear dynamic systems. Kybernetika, 2:331-346, 1966. Download.

V. Peterka and S. Bláha. Synthesis of regulation loops according to quadratic criterion. Kybernetika, 1:127-143, 1966. Download.

V. Peterka. New approach to identification of discrete systems. Kybernetika, 4:113-135, 1968. Download.

V. Peterka. Use of pseudo-random signals for identification of dynamic systems. Kybernetika, 5:406-421, 1969. Download.

A.V. Balakrishnan and V. Peterka. Identification in automatic control systems. pages 1-43. 1969. Download.

V. Peterka. Tally... In Preprints of the 2nd IFAC Symposium on Identification and Process Parameter Estimation, page paper ... Prague, 1970. Download.

V. Peterka. On steady-state minimum variance control strategy. Kybernetika, 8:219-231, 1972. Download.

V. Peterka. A square-root filter for real-time multivariable regression. Kybernetika, 11:53-67, 1975. Download.

V. Peterka. Experience accumulation for decision making in multivariate time series. volume 7, pages 143-159. Laxenburg, 1978. Download.

V. Peterka and M. Kárný. Bayesian system classification. In Preprints of the 5th IFAC Symposium on Identification and System Parameter Estimation, volume 1, pages 349-356. Darmstadt, 1979. Download.

V. Peterka. Bayesian system identification. In Preprints of the 5th IFAC Symposium on Identification and System Parameter Estimation, volume 1, pages 99-114. Darmstadt, 1979. Download.

V. Peterka. Real-time parameter estimation and output prediction for ARMA-type system models. Kybernetika, 17:526-533, 1981. Download.

V. Peterka. Bayesian approach to system identification. pages 1-80. 1981. Download.

J. Böhm and M. Kárný. Self-tuning regulators with restricted inputs. Kybernetika, 18(6):529-544, 1982. Download.

R. Kulhavý and M. Kárný. Tracking of slowly varying parameters by directional forgetting. In Preprints of the 9th IFAC World Congress, volume X, pages 178-183. Budapest, 1984. Download.

J. Böhm, A. Halousková, M. Kárný, and V. Peterka. Simple LQ self-tuning controllers. In Preprints of 9th IFAC World Congress, volume VII, pages 171-176. Budapest, 1984. Download.

V. Peterka. Algorithms for LQG self-tuning control based on input-output delta models. In Proc. of 2nd IFAC Workshop on Adaptive Systems in Control and Signal Processing, pages 13-18. Lund, Sweden, 1986. Download.

R. Kulhavý. Restricted exponential forgetting in real-time identification. Automatica, 23(5):589-600, 1987. Download.

V. Peterka. Self-tuning control with alternative sets of uncertain process models. In Proc. of IFAC Symposium on Adaptive Systems in Control and Signal Processing, pages 409-414. Glasgow, UK, 1989. Download.

R. Kulhavý and E. Kliokys. Tracking of time-varying parameters in delta models. Problems of Control and Information Theory, 28(2):107-123, 1989. Download.

E. Kliokys and R. Kulhavý. Bayesian filtering for discrete-time systems with random structure. In IFAC Symposium Adaptive Control and Signal Processing, pages 611-615. Glasgow, 1989. Download.

J. Böhm, M. Kárný, and R. Kulhavý. Practically-oriented LQ self-tuners. In Preprints of the IFAC Workshop on Evaluation of Adaptive Control Strategies in Industrial Applications. Tbilisi, 1989. Download.

V. Peterka. Predictive and LQG optimal control: equivalences, differences and improvements. In D.Henrichsen and B.Mårtenson, editors, Proc. of an Int. Workshop on Control of Uncertain Systems, pages 221 - 244. Birkhäuser, Boston - Basel - Berlin, 1990. Download.

This page contains algorithms published in different papers.

Sparse Blind Source Separation and Deconvolution

Orthogonal Variational PCA