Andrey Gavrilov

Activity Recognition Based on Semi-supervised Learning

Activity recognition is a hot topic in context-aware computing. In activity recognition, machine learning techniques have been widely applied to learn the activity models from labeled activity samples. Since labeling samples requires humans efforts, most existing research in activity recognition focus on refining learning techniques to utilize the costly labeled samples as effectively as possible. However, few of them consider using the costless unlabeled samples to boost learning performance. In this work, we propose a novel semi-supervised learning algorithm named En-Co-training to make use of the unlabeled samples. Our algorithm extends the co- training paradigm by using ensemble method. Experimental results show that En-Co-training is able to utilize the available unlabeled samples to enhance the performance of activity learning with a limited number of labeled samples.

In-building Localization using Neural Networks

Location awareness is key capability of context-aware ubiquitous environments. Received signal strength (RSS) based localization is increasingly popular choice especially for indoor scenarios after pervasive adoption of IEEE 802.11 wireless LAN. Fundamental requirement of such localization systems is to estimate location from RSS at a particular location. Multi-path propagation effects make RSS to fluctuate in unpredictable manner, introducing uncertainty in location estimation. Moreover, in real life situations RSS values are not available at some locations all the time making the problem more difficult. We employ modular multi-layer perceptron (MMLP) approach to effectively reduce the uncertainty in location estimation system. It provides better location estimation results than other approaches and systematically caters for unavailable signals at estimation time

A modular classification model for received signal strength based location systems

Estimating location of mobile devices based on received signal strength (RSS) patterns is an attractive method to realize indoor positioning systems. Accuracy of RSS based location estimation, particularly in large target sites, is effected by several environmental factors. Especially the temporal or permanent absence of radio signals introduces null values rendering sparsity and redundancy in feature space. We present a visibility matrix based modular classification model which systematically caters for unavailable signals. This model is practically realized using two eminent classification methods: (1) multi-layer perceptron and (2) LVQ. In order to confirm robustness and applicability of this model, we developed two location systems at different sites. Experimental results in real-world environments demonstrate that modular classification model consistently achieves superior location accuracy.

Modular Multilayer Perceptron for WLAN Based Localization

Location awareness is key capability of context-aware Ubiquitous environments. Received signal strength (RSS) based localization is increasingly popular choice especially for in-building scenarios after pervasive adoption of IEEE 802.11 wireless LAN. Fundamental requirement of such localization systems is to estimate location from RSS at a particular location. Multipath propagation effects make RSS to fluctuate in unpredictable manner, introducing uncertainty in location estimation. Moreover, in real life situations RSS values are not available at some locations all the time making the problem more difficult. We employ modular multi layer perceptron (MMLP) approach to effectively reduce the uncertainty in location estimation system. It provides better location estimation results than other approaches and systematically caters for unavailable signals at estimation time.

Context-aware, self-scaling Fuzzy ArtMap for received signal strength based location systems

Location awareness is the key capability of mobile computing applications. Despite high demand, indoor location technologies have not become truly ubiquitous mainly due to their requirements of costly infrastructure and dedicated hardware components. Received signal strength (RSS) based location systems are poised to realize economical ubiquity as well as sufficient accuracy for variety of applications. Nevertheless high resolution RSS based location awareness requires tedious sensor data collection and training of classifier which lengthens location system development life cycle. We present a rapid development approach based on online and incremental learning method which significantly reduces development time while providing competitive accuracy in comparison with other methods. ConSelFAM (Context-aware, Self-scaling Fuzzy ArtMap) extends the Fuzzy ArtMap neural network system. It enables on the fly expansion and reconstruction of location systems which is not possible in previous systems.

Using fuzzy decision tree to handle uncertainty in context deduction

In context-aware systems, one of the main challenges is how to tackle context uncertainty well, since perceived context always yields uncertainty and ambiguity with consequential effect on the performance of context-aware systems. We argue that uncertainty is mainly generated by two sources. One is sensors inherent inaccuracy and unreliability. The other source is deduction process from low-level context to high-level context. Decision tree is an appropriate candidate for reasoning. Its distinct merit is that once a decision tree has been constructed, it is simple to convert it into a set of human-understandable rules. So human can easily improve these rules. However, one inherent disadvantage of decision tree is that the use of crisp points makes the decision trees sensitive to noise. To overcome this problem, we propose an alternative method, fuzzy decision tree, based on fuzzy set theory.

devising a context selection-based reasoning engine for context-aware ubiquitous computing middleware

We propose a novel reasoning engine for context-aware ubiquitous computing middleware in this paper. Our reasoning engine supports both rulebased reasoning and machine learning reasoning. Our main contribution is to utilize feature selection method to filter the low-level contexts which are not useful for certain special high-level context reasoning. As a result, rules and learning models in the reasoning engines knowledge base are refined since useless context have been filtered. The merits of our proposed reasoning engine are described in details in this paper.

A Rapid Development Approach for Signal Strength Based Location Systems

Location systems are core technologies for enabling pervasive computing smart spaces. Signal strength based location estimation offers economical viability and sufficient accuracy for numerous location based services. Despite extensive research being carried out in this positioning method, tedious and lengthy development life cycle prohibits wide scale deployment of these systems. We present a new approach based on online machine learning paradigm which significantly reduces the development time. Besides rapid development of location systems this approach shows competitive location accuracy and allows incremental expansion of location system.Pervasive computing calls for applications which are often composed from independent and distributed components using facilities from the environment. This paradigm has evolved into task based computing where the application composition relies on explicit user task descriptions. The composition of applications has to be performed at run-time as the environment is dynamic and heterogeneous due to e.g., mobility of the user. An algorithm that decides on a component set and allocates it onto hosts accordingly to user task preferences and the platform constraints plays a central role in the application composition process. In this paper we will describe an algorithm for task-based application allocation. The algorithm uses micro-genetic approach and is characterized by a very low computational load and good convergence properties. We will compare the performance and the scalability of our algorithm with a straightforward evolutionary algorithm. Besides, we will outline a system for task-based computing where our algorithm is used.

Hybrid neural network model based on multi-layer perceptron and adaptive resonance theory

The model of the hybrid neural network is considered. This model consists of model ART-2 for clustering and perceptron for preprocessing of images. The perceptron provides invariant recognition of objects. This model can be used in mobile robots for recognition of new objects or scenes in sight the robot during his movement.

Usage of Hybrid Neural Network Model MLP-ART for Navigation of Mobile Robot

We suggest to apply the hybrid neural network based on multi layer perceptron (MLP) and adaptive resonance theory (ART-2) for solving of navigation task of mobile robots. This approach provides semi supervised learning in unknown environment with incremental learning inherent to ART and capability of adaptation to transformation of images inherent to MLP. Proposed approach is evaluated in experiments with program model of robot.