Petia Koprinkova-Hristova

Artificial Neural Networks and Machine Learning – ICANN 2013

When the inputs of a regression problem are corrupted with noise, integrating out the noise process leads to biased estimates. We present a method that corrects the bias caused by the integration. The correction is proportional to the Hessian of the learned model and to the variance of the input noise. The method works for arbitrary regression models, the only requirement is two times differentiability of the respective model. The conducted experiments suggest that significant improvement can be gained using the proposed method. Nevertheless, experiments on high dimensional data highlight the limitations of the algorithm.

New Method for Nonlinear Fuzzy Correction Modelling of Dynamic Objects

In the paper a method to use the equivalent linearization technique of the nonlinear state equation with the coefficients generated by the fuzzy rules for current operating point is proposed. On the basis of the evolutionary strategy and properly defined identification procedure, the fuzzy rules are automatically designed to maximize the accuracy of the resulting linear model.

Adaptive Critic Design with Echo State Network

In the present paper an application of a novel neural network architecture called Echo State Network (ESN) within the frame of a reinforcement learning scheme named Adaptive Critic Design (ACD) is proposed. Our aim is to investigate the possibility for on-line training of adaptive critic using the ESN architecture. In particular the application of this approach to mobile robot control is presented. Our preliminary results are encouraging and demonstrate that ESNs are good candidates for the on-line application of an ACD optimization approach due to their specific structure and fast training algorithm.

Echo state networks for multi-dimensional data clustering

In the present work we showed that together with improved stability the Intrinsic Plasticity (IP) tuned Echo State Network (ESN) reservoirs possess also better clustering abilities that opens a possibility for application of ESNs in multidimensional data clustering. The revealed ability of ESNs is demonstrated first on an artificially created data set with known in advance number and position of clusters. Automated procedure for multidimensional data clustering was proposed. It allows discovering multidimensional data structure without specification in advance the clusters number. The developed procedure was further applied to a real data set containing concentrations of three alloying elements in numerous steel compositions. The obtained number and position of clusters showed logical from the practical point of view data separation.

BACKPROPAGATION THROUGH TIME TRAINING OF A NEURO-FUZZY CONTROLLER

The paper considers gradient training of fuzzy logic controller (FLC) presented in the form of neural network structure. The proposed neuro-fuzzy structure allows keeping linguistic meaning of fuzzy rule base. Its main adjustable parameters are shape determining parameters of the linguistic variables fuzzy values as well as that of the used as intersection operator parameterized T-norm. The backpropagation through time method was applied to train neuro-FLC for a highly non-linear plant (a biotechnological process). The obtained results are discussed with respect to adjustable parameters rationality. Conclusions are made with respect to the appropriate intersection operations too.

ESN intrinsic plasticity versus reservoir stability

The work presented in this paper was inspired by similarities between intrinsic plasticity (IP) pre-training of the ESN reservoir and the common RNN stability conditions derived from nonlinear control theory. The common theoretical stability conditions were applied to the ESN structure. It was proven that in fact IP training achieves a balance between maximization of entropy at the ESN output and the concentration of that output distribution around the pre-specified mean value. Thus the squeezing of the neuron nonlinearities is produced not only by nonzero biases and translation of the ESN equilibrium state but also by the chosen output distribution mean value. The numerical investigations of different random reservoirs showed that the IP improvement stabilizes even initially unstable reservoirs.

Adaptive critic design with ESN critic for bioprocess optimization

We propose an on-line action-dependent heuristic dynamic programming approach based on recurrent neural network architecture - Echo state network (ESN) - as critic network within the frame of adaptive critic design (ACD), to be used for adaptive control. Here it is applied to the optimization of a complex nonlinear process for production of a biodegradable polymer, briefly called PHB. The on-line procedure for simultaneous critic training and process optimization is tested in the absence and presence of measurement noise. In both cases the optimization procedure succeeded in increasing the productivity and in proper training of the adaptive critic network at the same time.

A New Method for Generating of Fuzzy Rules for the Nonlinear Modelling Based on Semantic Genetic Programming

In this paper we propose a new approach for nonlinear modelling. It uses capabilities of the Takagi-Sugeno neuro-fuzzy systems and population based algorithms. The aim of our method is to ensure that created model achieves appropriate accuracy and is as compact as possible. In order to obtain this aim we incorporate semantic information about created fuzzy rules into process of evolution. Our method was tested with the use of well-known benchmarks from the literature.

New Method for Non-linear Correction Modelling of Dynamic Objects with Genetic Programming

In the paper a method to adapt the equivalent linearization technique of the non-linear state equation is proposed. This algorithm uses correction matrices. It also uses arrays amendments which elements are determined for each new point. These elements are generated by a formula created automatically using genetic programming.

Sound fields clusterization via neural networks

Paper presents application of a recently proposed approach for multidimensional data clustering to data received from a microphone array antenna. The accumulated sound pressure at each point (a microphone in the array) is used to create “sound picture” of the observed by the microphone antenna area. Features for classification are extracted using overlapping receptive fields based on the model of direction selective cells in the middle temporal (MT) cortex. Next the clustering procedure using Echo state network and subtractive clustering algorithm is applied to separate receptive fields in proper number of classes. The obtained results are compared with the sonograms created by the original software of the producer of microphone array.