Catálogo de publicaciones - libros

The article describes multi-agent engine for data clustering and IF-THEN rules generation and their application to transportation logistics. The developed engine can be used for investigating customer source data, pattern discovery in batch or in real time mode and ongoing forecasting and consolidation of orders and in other cases. Engine basic architecture fits well for both batch and real time clustering. The example of data clustering and generation of IF-THEN rules for one of UK logistics operators is considered. It is shown how the extracted rules were applied to automatic schedule generation and how as a result the quality of schedules was improved. The article also describes an approach, which allows getting orders consolidation from extracted rules. Algorithm of rule search and the obtained results analysis are other points mentioned.

Over the last years, the data generated in Telecommunications Networks has reached unmanageable limits of information. Data Mining (DM) techniques have showed their advantages on helping to manage this information and transforming it in useful knowledge. However, due to the dynamics of the environment of Telecommunications Networks, the simple application or adaptation of DM techniques is not enough to obtain timely a deeper knowledge. In this paper, this problem is addressed by applying DM techniques in real time. First, we propose a methodology taking into account all the processes involved in transforming telecommunications data into information, and finally to knowledge. Second, we propose a framework for the utilization of Intelligent Agents to help the process of DM in real time. To illustrate our approach, we describe a real-life case study based on the integration of Intelligent Agents and DM technologies for obtaining in real time knowledge that is critical for managing telecommunication networks.

Distributed decision making and learning procedures operate with distributed data sources. Intelligent access to distributed data sources is one of the important requirements to any distributed learning and decision making systems. In many such applications, input data arrive from spatially distributed sensors structured in sensor networks. An increasing interest to what is called for which agent-based technology seems to be rather attractive implementation paradigm forms a marked trend in the sensor network area. Indeed, the latter provides a natural mapping from the set of nodes of intelligent sensor network to the set of collaborating agents and, at the same time, enriches the network nodes by the methods of collaboration developed under multi-agent approach and provides the network nodes with the capability to communicate in terms of high level language. The paper proposes a typical (reusable) architecture of the intelligent sensor agents as well as typical protocols supporting interaction of sensors with other software components of agent-based network intended for distributed learning and classification. The main solutions proposed in the paper are demonstrated via prototyping of two case studies.

A framework for deploying and coordinating agents for data mining over a distributed system by using the notion of dynamics between components is presented. The key idea behind the framework is to provide agents with deployment policies for their deployment and coordinations as relations between them and other agents. As a result, a federation of distributed agents can migrate and coordinate over a distributed system in a self-organizing manner. This paper also presents a prototype implementation of the approach and its applications.

Emergency response operations require availability of tools that would allow fast and clear description of situation, generation of effective solutions for situation management, selection of a right decision maker and supplying him/her necessary data. During decision making support a large amount of raw data describing current situation, users preferences, found solutions and final decisions done by decision maker are accumulated in the repository. To make these data useful, methods of data mining can be applied. The goal of decision mining is to find “rules” explaining under which circumstances one activity is to be selected rather than the other one. The paper presents results of a research concerning mining of decisions stored in user profiles to find common preferences for different roles of decision makers participating in emergency response operations. These preferences could be used as a basis for building of decision trees allowing (semi)automatically selection of best decisions in a typical situation. Modeling of an emergency response management system implementing the research results has been done using software agents playing roles of different types of the system users. Case studies related to fire and accident response operations have been used.

As shopbots spread through Internet collecting information about lowest prices / highest qualities human sellers will turn out to be too slow to tune the prices and thus unprofitable in comparison with smart agents — pricebots. One of the most promising approaches to building pricebots is -learning. Its advantages: flexibility to act under changing conditions of virtual markets, -learning sellers can take into account not only immediate rewards but also profits far ahead, and don’t need information neither on buyer demand nor on competitors’ behaviour. But up to now -learning sellers used state representation exponential in the number of sellers acting in the market and could function successfully only with one competitor which no doubt is unrealistic. We are proposing a new state representation independent of the number of sellers that allowed to 10 agents to find the prices that maximize cumulative profits under conditions of high competition in three moderately realistic economic models. It was also shown that due to their flexibility Q-learning sellers managed to collect more profit than pricebots based on two other generally used approaches even though one of them possessed much more information about buyer demand and competitors’ behaviour. The proposed representation doesn’t depend on the number of sellers and in principle -learning sellers using it can function in the markets with arbitrary number of competitors.

In a multi-agent system, agents need to identify their relationships in order to coordinate their actions. Actually, coordination is the most challenging issue in multi-agent design. Very often, the definition of agents’ relationships depends on the properties of the tasks they are able to perform. Knowing its task, an agent is able to build a dependence network that helps identify other agents with whom to interact and cooperate. An agent can also fuse branches in its dependence network in order to derive dependence relations between other agents. The fusion of dependence relations is a powerful tool that could be used by agents to coordinate their tasks and enhance the whole system performance.

The paper proposes the framework for investigation of prospective adaptive and cooperative defense mechanisms against the Internet attacks. The approach suggested is based on the multi-agent modeling and simulation. According to the approach the defense and attack systems are represented as interacting teams of intelligent agents that act under some adaptation criterion. They adjust their configuration and behavior in compliance with the network conditions and attack (defense) severity. The paper represents the architecture and software implementation of simulation environment that combines discrete-event simulation, multi-agent approach and packet-level simulation of various Internet protocols. The environment allows to simulate complex attack and defense scenarios. The paper describes the experiments aimed on the investigation of adaptive “Distributed Denial of Service” attacks and defense mechanisms.

We aim to advance in constructing collaborative agents able to acquire the contents of human vocabulary associated with competitions. Refining the framework and criteria of performance of agents we project the study on the class of game tree represented competition problems. For known representative of the class – chess like combinatorial games, we categorize the contents of a comprehensive repository of units of chess vocabulary by formal structures of attributes, goals, strategies, plans, etc. We define Personalized Planning and Integrated Testing algorithms able to elaborate moves in target positions dependent on those categories of chess knowledge. We then demonstrate the effectiveness of the algorithms by experiments in acquisition the solutions of two top Botvinnik’s tests – the Reti and Nodareishvili etudes. For min max game tree based search algorithms these etudes appears to be computationally hard due the depth of the required analysis and very dependence on the expert knowledge.

This paper describes a possibility of applying on-line learning techniques to train agents for teamwork. Special modules proposed based on immunological networks capable of on-line learning in dynamically changing environments. The above modules provide adaptive agents’ behavior for teamwork after they are trained to select of primitive behaviors under variable environmental conditions. Reinforcement learning is considered to be the main method for training agents during a game. The special agent capable of on-line training for basketball competitions in RoboFIBA virtual system was developed and investigated. Examples of immunological networks for agent teamwork implementation are considered, and results of experiment with hard and training agents are described.