Dragan Jevtic

Predicting user movement for advanced location-aware services

-The paper deals with prediction of user movement in the context of enhancing location-aware services. Location and movement information are based on the simplest mechanism provided by mobile networks -broadcasted cell identification. The proposed approach includes the following steps: collecting information about user movement, analysing and learning user movement patterns, and applying movement knowledge to movement prediction. Movement prediction system based on the neural networks used for movement regularity detection and movement prediction is presented. Service architecture and an example service concludes the paper.

Location name extraction for user created digital content services

The increase in the amount of electronically stored textual data over the past decade has opened new communication possibilities in which users significantly participate in services. The result is a variety of digital content services in which new formats, such as free text form, have become particularly accepted and attractive. To support user service interaction, services must be able to obtain specific information from user defined texts. This paper proposes a model for location name extraction, constructed using a neural network trained with a Backpropagation learning algorithm. An analysis of location name interpretation, the semantic and binding problem in Croatian, as well as language specific variations are given.

The Intelligent Agent-Based Control of Service Processing Capacity

The paper presents new approach to processing capacity protection in the service system with multiple server units. In the integrated service communication networks an important problem is to implement call admission control and routing so as to optimally use the network resources. We assumed reality of several classes of jobs and tried to preserve an amount of processing capacity for high priority jobs that can arrive in a burst. In parallel, current low priority jobs were processed with continuous regulation of servers load. Simulation results showed rapid adaptation and good balancing around the predetermined maximum processing level. Inspiration was found in the paradigm of software agent technology and potential advantages appearing when applied in telecommunication network. Main characteristic of the usage of intelligent agent is the opportunity to permanently transfer adaptation regarding one or more parameters following optimal and requested policy.

User movement prediction based on traffic topology for value added services

Value added services are based on user context awareness. Important context aspect is location, which could be extended to future locations if services had the ability to predict movement. We propose a model for user movement prediction based on traffic topology. Benefits of the model are presented on example service, while the performance is evaluated on real user movement data.

User privacy risk calculator

User privacy is becoming an issue on the Internet due to common data breaches and various security threats. Services tend to require private user data in order to provide more personalized content and users are typically unaware of potential risks to their privacy. This paper proposes a risk calculator based on a feedforward neural network that will provide users with an ability to calculate risks to their privacy. The proposed calculator is evaluated on a set of real world example scenarios. Furthermore, to give more insight into privacy issues, each estimated risk is explained by several real life scenarios that might happen if the observed parameters are obtained by an attacker. In turn, this should raise user awareness and knowledge about privacy issues on the Internet.

The effect of alteration in service environments with distributed intelligent agents

The paper presents some properties of the communication network supported by the intelligent agents. The intelligent agents were placed into network nodes and they were immobile. They were used to regulate the transfer of mobile agents from the network input, through the routing nodes and finally, towards the service processing nodes. These nodes were the programs running on the computers. A new model of distributed and collaborating intelligent agents was designed and presented. Continuous adaptation and agents collaboration was achieved by reinforcement Q-learning. For such a model the results show rapid tendency to reduce state time of mobile agents in the service. The obstructive effects expansion to the other agents when a change of processing capabilities in the region occurs was detected and described.

Self-trained agents optimize communication service by intelligent selection

The paper presents a method using an optimal selection of an agent in a thought client-server environment. The selection criteria are based on continuous learning and monitoring of the agents behavior. The work has been motivated by the different abilities and properties of the agents in the network, particularly when they act in distributed environments. The main idea presented is a permanent transfer adaptation of the requests from a client agent to an optimal service provider agent. Continuous adaptation is achieved by reinforcement Q-learning. Simulation results show that by implementing knowledge into an agents behavior, it is possible, in particular situations, to significantly accelerate the service system.

Estimating real world privacy risk scenarios

User privacy is becoming an issue on the Internet due to common data breaches and various security threats. Services tend to require private user data in order to provide more personalized content and users are typically unaware of potential risks to their privacy. This paper continues our work on the proposed user privacy risk calculator based on a feedforward neural network. Along with risk estimation, we provide the users with real world example scenarios that depict privacy threats according to selected input parameters. In this paper, we present a model for selecting the most probable real world scenario, presented as a comic, and thus avoid overwhelming the user with lots of information that he/she may find confusing. Most probable scenario estimations are performed by artificial neural network that is trained with real world scenarios and estimated probabilities from real world occurrences. Additionally, we group real world scenarios into categories that are presented to the user as further reading regarding privacy risks.

Load protection model based on intelligent agent regulation

Paralleling rapid advancement in the telecommunication network expansion is necessary for advanced network traffic management surveillance. The increasing number and variety of services being offered by networks have emphasized the demand for optimized load management strategies. The paper deals with regulation of a mobile agent moving toward service processing resource in the part of the agent network. We have constructed the agent architecture for the control of service processing load. The goals of the controlling system were both to protect the processing load and to predict the arrival rate of clients requests. Self-adaptive property is implemented by reinforcement Q-learning. The analysis is based on experimentation through simulations.

Intelligent call transfer based on reinforcement learning

This paper presents an application of reinforcement Q-learning in solving the problem of an automatic call transfer. Given outcomes are the results of simulations. A potential problem has been detected in a telecommunication call center in which a particular association of agents is applied to serve large number of the incoming calls. The main idea presented here is the adaptation of the current call transfer toward an optimal agent. The search for an optimal agent is based on its current and previous activity. The simulations show that implementation of the knowledge about the agent behavior can, in particular situations, significantly accelerate the service system.