DonNTU   Masters' portal

Neural models of software agents in socially-oriented multi-agent systems

Content

Introduction

In the science of "artificial intelligence" (AI), until recently dominated the formal (rationalist) approach based on classical logic, the symbolic paradigm and traditional approaches in knowledge engineering. Now in the AI start to develop new directions: semiotic model, "soft computing" and "computational intelligence", multi-agent systems (MAS) and artificial organization, which allows you to create intelligent systems of new generation. Great perspective on this evolutionary path of development of AI associated with the ideas and principles of synergy in AI, meaning joint action and cooperation [5]. The principle of harmonization of individual interests and different points of view, characteristic of groups and organizations are increasingly beginning to be used by practitioners of AI. It was taken as the basis for the design of artificial agents interactions, building a MAS and intellectual organizations.

G.Saymon violated the established tradition of rationalism in computer science and AI. He formulated the important principle of bounded rationality [4]. According to this principle, there are limits to the rationality of decision-makers due to their individual psycho-physiological constraints (speed of information processing, memory, etc.). Decision-making in organizations is always in the face of uncertainty, risk, ie in circumstances where the manager can not in principle cover all alternatives, and evaluate the consequences of decisions. Thus, the behavior of a separate, stand-alone agent in principle can not reach an absolute (or even higher) level of rationality. The number of alternatives to be considered, usually so great, and the information needed to evaluate them so vast that even an approximation to the absolute rationality is impossible. Therefore, management decisions in a large organization, understood as a network of agents that are not based on full information and optimization, and on meeting the conflicting criteria of various agents (nodes) in the face of uncertainty.

This raises the important task of analysis and research on methodological recommendations arising from the theory of bounded rationality, which will help in the development of intelligent multi-agent type organizations. The main question addressed in this paper - to find a way to best describe the situation and the decision-making agent, which are responsible for taking decisions.

1. The purpose and objectives of the study, expected results

The purpose of the final work is to study multi-agent systems, software agents and their applicability to the task, and the creation of software with multi-agent system MadKit.

Research Objectives:

  1. Review and development of instrumental multi-agent systems development environments.
  2. Modeling and architecture of the developed system of agents.
  3. Selection methods of interacting agents.
  4. Software implementation of multi-agent system.
  5. Analysis of simulation results.

Object of research: modeling of multi-agent systems.

Expected results are as follows:

  1. Development of a new scientific field of multi-agent systems
  2. Improving the quality of modeling socio-economic processes through the use of software agents.
  3. Analysis of simulation results, evaluation of the approximate model to reality.

2. Relevance of the topic

Multiagent Systems - one of the new paradigm in artificial intelligence, which is used to solve these problems that are difficult or impossible to solve with a single agent or monolithic system. Examples of these tasks are on-line commerce, disaster management, and modeling of social and economic structures.

Multiagent systems are already being applied in our lives, such as graphics-intensive applications, movies. The theory of the MAS is used in composite systems, defense, transport, logistics and many others. Multiagent systems as proven in the field of network and mobile technologies.

Thus, multiagent systems are an effective tool for solving complex problems in a large range of subject areas.

3. The scientific novelty

As a scientific novelty of this work are the models and algorithms developed by software agents that can bring socio-economic modeling of the processes to a new level.

4. A review of research and development on the subject

The theme of multi-agent systems interested in Donetsk National Technical University, conducted research and development in this field and presents the works of masters: Lukina Yulia, Zudikova Julia, Evdokimov Andrey, Lyamin Roman and others.

issues of agent-oriented programming is actively engaged scientists and programmers around the world, proof of this is one of the largest web sites for placement of electronic scientific publications http://dl.acm.org/ at the request of "multi-agent system" issuing about 29 thousand different publications and articles. Similar results (20 thousand publications) provides web site http://www.springerlink.com/

You can also note that the worldwide multi-agent systems are used in various fields of human activity, such as education, medicine and economics.

5. Rationalist paradigm in computer science and artificial intelligence

This paradigm has until recently not only in fundamental computer science and artificial intelligence, but also in related disciplines: management theory, decision theory, cognitive psychology. Faithful companion of rationalism is the principle of reductionism - the real information of the complex phenomenon of highly simplified models. It is believed that the decision of the agent is rational if his choice of options can give the most favorable results, taking into account all the constraints.

Prerequisites are rational solutions [9] :
1) clarity problem;
2) focus on the goal;
3) knowledge of all the options and their consequences are known;
4) the clarity of the benefits, ie the benefits are clear;
5) The constancy of the benefits;
6) lack of time or physical limitations;
7) maximum output, ie final selection will maximize the benefits of this choice.

Fully rational agent (born rational agent) - an agent acting optimally to achieve the best expected outcome. This term is one of the most fundamental in the economy, decision theory and artificial intelligence [6] . The concept of "rational agent" in AI has come out of the economy, and committed him to the real revolution, bringing together disparate areas of research.

Because in the real world the absolute (full) rationality is almost inaccessible, it is expected that the agent in the decision-making is constrained by the principle of bounded rationality.

6. The basic principles of the theory of bounded rationality

A recognized pioneer generalized model of economic behavior, which was called the theory of bounded rationality is considered the Nobel Prize-winning American economist, professor of psychology and computer science G.Saymon [5] .

Picture 1 Algorithm model built on the concept of bounded rationality

Picture 1 - Algorithm model built on the concept of bounded rationality

His theory was based on of the following observations:
1. Absolute Rationality requires a complete knowledge of all the decisions. In reality, this is not achievable.
2. There are fundamental limitations on the ability to forecast the consequences of decisions (as to future events).
3. Absolute Rationality requires a choice of all possible behaviors. In practice, only a small number of options can be taken into account.


G.Saymon and his colleagues conducted a number of empirical studies of how real is the process of decision-making in firms and, based on regulations developed algorithms for making the "right" decisions [1] .

In order to maximize utility or profit, the economic entity is simply not enough computing ability. The problem of the subject is not so much that he has little information as to the fact that it is too much on the capacity of its processing. The process of decision-making model of G.Saymon can be described by two main concepts - search and making a satisfactory option.

According G.Saymon may not be comprehensive utility function, which would compare the diverse alternatives. This feature, in his opinion, has only two {0, 1} or three {-1, 0, 1} values, where 1 denotes a satisfactory option, -1 - poor, and 0 - indifferent.

As a result of the economic subject of searches for solutions to the problem as long as it finds the first acceptable solution, ie comes as shown in Figure 1.

Acceptability or otherwise of each option determines for himself. G.Saymon describes the process using borrowed from psychology category of "level of aspiration." The concept of level of aspiration suggests that at one time a person has some idea of what it may expect. The level of claims is not frozen, the bar moves all the time depending on the results of the last step. If he is successful, the level of claims goes up - a man set himself a higher goal. In case of failure, the level of claims down, as a person begins a more critical attitude to their abilities. Option is considered to be satisfactory, if it allows a person to overcome the bar - the level of claims.

Can be seen that the choice of a satisfactory option requires that the entity is much less awareness of art and countable. He is no longer necessary to have accurate information about the outcome of this case and compare it with the outcomes of alternatives within the overall utility function. Suffice intuitive idea of what this option is above or below an acceptable level. At the same time to compare options with each other generally do not need to [1] .

Concept of bounded rationality is the only formal model of the economic theory of human behavior, an alternative utility maximization and profit, although its application in practice requires complex formulas and calculations. At present the model of "bounded rationality" has been used successfully in regulatory guidelines, and computer programs.

7. Using the theory of bounded rationality in multiagent systems

Concept of bounded rationality has been widely used in the theory of multi-agent systems. MAS is composed of several interacting agents, which may have a different architecture. Depending on the type of agents, MAS can be used to solve many problems that are difficult or impossible to solve with a single agent or monolithic system. In Picture 2 shows one of the intelligent agent architectures. It consists of four components. The most significant difference is observed between the training component, which is responsible for making improvements, and productive component that provides a choice of external actions [7] .

Productive component gets out of the perceived environmental information and decides on the response. Training component uses feedback information from the critique of the evaluation of how the agent determines how to be a productive component is modified so that it has successfully operated in the future.

Picture 2 Architecture of an intellectual agent

Picture 2 - Architecture of an intelligent agent

(animation - Resolution: 450 x 337 px; size: 25 kb; frame 9, the delay between shots: 0.5 s, the number of iterations: 5)

Critic says the training component of how well the agent is given the prescribed standard of performance. The critic is necessary because the perception of the results themselves do not give any indication as to whether the success of the agent.

The final component of the agent is a generator of problems. His task is to propose actions that should result in a new and informative experience [8] .

Conclusion

At this stage of research the authors focus on the analysis of the concept of bounded rationality in human behavior considered from the perspective of artificial intelligence.

According to experts in the field of classical theory of human behavior, based on the principle of full rationality in decision-making man, far from reality. The factors limiting the rationality of agents include:
1) lack of motivation;
2) The local nature of the available information (resource constraints);
3) the fundamental incompleteness and inaccuracy of the information received from other agents;
4) the effects of random environment.

Any agent in the organization is always in interaction with other agents, negotiating, trying to create a coalition, and so on, so his perception of the subjective and fragmentary, and the knowledge needed to solve the problem, vague, inaccurate and limited to [5] . So now, in ICA is dominated by the thesis of the desirability of such conditions using collective decision-making agents in the course of their joint activities. On the basis of this thesis built a new behavioral theory of organizations. It is based on recognition of the limitations of pure rationalism and the relative attraction of the ideas of rationality in human behavior, intellectualization and biological function of organizations. Therefore, if the MAC should work in a complex, transformed, dynamic environment, characterized by a high level of uncertainty, it must be based on these new principles:
1) instead of a rigid "tree" hierarchy, it is expedient to use a "flat" network structure, in which agents are not seen as a gear, but as nodes in the network, realizing goals, and implement the designed MAC interaction with other nodes [5 ] ;
2) to provide intelligent agents are not narrow, but wider field of expertise, providing the solution of large problems;
3) to move from one-man to the cooperation and coordination.


This approach to building multi-agent models of heterogeneous dynamical systems with distributed intelligence will more realistically simulate the entire system, as well as various aspects of human behavior, participants in such systems.

Note

In writing this essay master's work is not yet complete. Final completion: December 2012. The full text of the and materials on the topic can be obtained from the author or his head after that date.

References

  1. Автономов В. История экономических учений/Под ред. О.Ананьина, Н. Макашева:Учеб. пособие. — М: ИНФРА-М, 2002. - 784 с. — (Серия «Высшее образование»).
  2. Адылова Ф.Т. Технология грид и мультиагентных систем: новая парадигма развития телемедицины [Электронный ресурс]. – Режим доступа: http://www.nbuv.gov.ua/portal/Chem_Biol/Ujtm/2008_2/2008_2_8.pdf
  3. Меркулова Т.В., Акулов Н.В. Мультиагентный подход к моделированию экономической системы в условиях нестационарной внешней среды [Электронный ресурс]. – Режим доступа: http://www.nbuv.gov.ua/Portal/Soc_Gum/Bi/2009_4/4(2)/117-119.pdf
  4. Саймон Г. Науки об искусственном: Пер. с англ. – М.: Мир, 1972.
  5. Тарасов В.Б. От многоагентных систем к интеллектуальным организациям:философия, психология, информатика.- М.: Эдиториал УРСС, 2002. – 352 с.
  6. Рациональный агент. [Электронный ресурс]. – Режим доступа: http://ru.wikipedia.org/wiki/Рациональный_агент
  7. Многоагентная система. [Электронный ресурс]. – Режим доступа: http://ru.wikipedia.org/wiki/Многоагентная_система
  8. Интеллектуальные агенты. [Электронный ресурс]. – Режим доступа:http://www.itfru.ru/index.php/intellectual-agents
  9. Принятие решений. [Электронный ресурс]. – Режим доступа:http://club-energy.ru/8_2.php
  10. Лямин Р.В. Многоагентная система обучения студентов на кафедральном уровне [Электронный ресурс]. – Режим доступа: http://masters.donntu.ru/2008/fvti/lyamin/diss/index.htm