JDJakub Dokoupil

Jakub Dokoupil

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Jakub Dokoupil has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Bayesian Knowledge Transfer for a Kalman Fixed-Lag Interval Smoother

    by OndřEj Skalský, Jakub Dokoupil

    A Bayesian knowledge transfer mechanism that leverages external information to improve the performance of the Kalman fixed-lag interval smoother (FLIS) is proposed. Exact knowledge of the external observation model is assumed to be missing, which hinders the direct application of Bayes' rule in traditional transfer learning approaches. This limitation is overcome by the fully probabilistic design, conditioning the targeted task of state estimation on external information. To mitigate the negative impact of inaccurate external data while leveraging precise information, a latent variable is introduced. Favorably, in contrast to a filter, FLIS retrospectively refines past decisions up to a fixed time horizon, reducing the accumulation of estimation error and consequently improving the performance of state inference. Simulations indicate that the proposed algorithm better exploits precise external knowledge compared to a similar technique and achieves comparable results when the information is imprecise.

  • Open Access

    Bayesian Knowledge Transfer for a Kalman Fixed-Lag Interval Smoother

    by Ondřej Skalský, Jakub Dokoupil

    A Bayesian knowledge transfer mechanism that leverages external information to improve the performance of the Kalman fixed-lag interval smoother (FLIS) is proposed. Exact knowledge of the external observation model is assumed to be missing, which hinders the direct application of Bayes’ rule in traditional transfer learning approaches. This limitation is overcome by the fully probabilistic design, conditioning the targeted task of state estimation on external information. To mitigate the negative impact of inaccurate external data while leveraging precise information, a latent variable is introduced. Favorably, in contrast to a filter, FLIS retrospectively refines past decisions up to a fixed time horizon, reducing the accumulation of estimation error and consequently improving the performance of state inference. Simulations indicate that the proposed algorithm better exploits precise external knowledge compared to a similar technique and achieves comparable results when the information is imprecise.


  • Recursive Identification of the ARARX Model Based on the Variational Bayes Method

    by Jakub Dokoupil, Pavel Václavek

    Bayesian parameter estimation of autoregressive (AR) with exogenous input (X) systems in the presence of colored model noise is addressed. The stochastic system under consideration is driven by colored noise that arises from passing an initially white noise through an AR filter. Owing to the additional AR filter, the ARARX schema provides more flexibility than the ARX one. The gained flexibility is countered by the fact that the ARARX system is no longer linear-in-parameters unless the white noise components or the AR noise filter are available. This paper analyzes the problem of estimating the unknown coefficients of the ARARX system and the model noise precision under conditions where the AR noise filter is both available and unavailable. While the former condition reduces the estimation problem to standard linear least squares, the latter one gives rise to an analytically intractable estimation problem. The intractability is resolved by the distributional approximation technique based on the variational Bayes (VB) method.


  • Open Access

    Recursive Identification of Time-Varying Hammerstein Systems With Matrix Forgetting

    by Jakub Dokoupil, Pavel Václavek

    The real-time estimation of the time-varying Hammerstein system by using a noniterative learning schema is considered and extended to incorporate a matrix forgetting factor. The estimation is cast in a variational-Bayes framework to best emulate the original posterior distribution of the parameters within the set of distributions with feasible moments. The recursive concept we propose approximates the exact posterior comprising undistorted information about the estimated parameters. In many practical settings, the incomplete model of parameter variations is compensated by forgetting of obsolete information. As a rule, the forgetting operation is initiated by the inclusion of an appropriate prediction alternative into the time update. It is shown that the careful formulation of the prediction alternative, which relies on Bayesian conditioning, results in partial forgetting. This article inspects two options with respect to the order of the conditioning in the posterior, which proves vital in the successful localization of the source of inconsistency in the data-generating process. The geometric mean of the discussed alternatives then modifies recursive learning through the matrix forgetting factor. We adopt the decision-making approach to revisit the posterior uncertainty by dynamically allocating the probability to each of the prediction alternatives to be combined.


  • Variable Exponential Forgetting for Estimation of the Statistics of the Normal-Wishart Distribution with a Constant Precision

    by Jakub Dokoupil, Pavel Vaclavek

    The problem of estimating normal regression-type models with possibly time-varying regression parameters and constant noise precision is considered and examined from the Bayesian viewpoint. The solution we propose exploits a collaborative decision in order to face the incomplete model of parameter variations. Under this approach, a loss functional evaluating two prediction alternatives is constructed, which allows us to merge both alternatives, complying with the principles of optimization theory. Specifically, the posterior probability density function (pdf) and its flattened variant are combined by means of the geometric mean with automatically adjusted weights. The result is an automatic rescaling of the covariance matrix through the forgetting factor in response to empirically confirmed performance.


  • Forgetting Factor Kalman Filter with Dependent Noise Processes

    by Jakub Dokoupil, Pavel Vaclavek

    The paper addresses the problem of filtering the state of a normal dynamical process with a dependency between the process and the measurement noise variables in the presence of an inaccurate model description. As regards the time occurrence of the noise dependency, we discuss the dependency structure where both the variables are correlated at the same time. An adaptive formulation of the Kalman filter (KF) is designed in order to mitigate the impact of the process model uncertainty on the degradation of the filter performance. The filter we propose exploits the collaborative decision to introduce a variable forgetting factor into the time update to reduce artificially the effect of obsolete knowledge on the filtering solution. Within the decision-making rules, a loss functional quantifying the time update is constructed to optimally combine the prediction alternatives possessing the form of the normal probability density function (pdf). The result is an adjustment of the Kalman gain matrix in response to empirically confirmed performance.


  • Data-Driven Stabilized Forgetting Design Using the Geometric Mean of Normal Probability Densities

    by Jakub Dokoupil, Pavel Vaclavek

    This paper contributes to the solution of adaptive tracking issues adopting Bayesian principles. The incomplete model of parameter variations is substituted by relaying on the use of data-suppressing procedure with two goals pursued: to provide automatic memory scheduling through the data-driven forgetting factor, and to compensate for the potential loss of persistency. The solution we propose is the geometric mean of the posterior probability density function (pdf) and its proper alternative, which, for the normal distribution, can be reduced to the convex combination of the information matrix and its regular counterpart. This coupling policy results from maximin decision-making, where the Kullback-Leibler divergence (KLD) occurs as a measure of discrepancy. In this context, the weight (probability) assigned to the information matrix is regarded as the forgetting factor and is controlled by a globally convergent Newton algorithm.


  • Regularized Extended Estimation With Stabilized Exponential Forgetting

    by Jakub Dokoupil, Alina Voda, Pavel Vaclavek

    This technical note concerns the problem of variable regularized estimation of time-varying nonlinear systems from the Bayesian viewpoint. The questions of how to impose the posterior information being variably regularized and how to forget this information are carefully discussed. The estimator design adopts the strategy of the iterated Kalman filter but differs in that, instead of the separated moments of the linearized system, only the augmented covariance matrix is updated. To suppress obsolete information, a decision problem involving the Kullback–Leibler divergence is solved. The decision provides the best combination of a pair of time-evolution model hypotheses in terms of the geometric mean. As a result, exponential forgetting with the adaptively tuned factor is inserted into the estimation process. The regularization of the investigated statistics is induced through the processing of externally supplied information. The presented estimator allows for absolute discarding or, conversely, retention of external information produced in terms of the Normal–Wishart distribution.