Various intelligent agents are already being used and studied in the world. Sometimes people may not even realize that they are using smart agents in their life. In our work, we investigated monitoring agents which task is to transform information. One of the areas of use of monitoring agents is the financial exchange, which makes this work interesting for a wide range of people.
The use of an agent approach to build
monitoring information systems (MIS) is the basis
of the concept of intelligent monitoring [
In modern world, intelligent agents are widely
used in various industries, social and political
spheres. Agent models are a powerful tool for
studying the object of monitoring, which allows
to describe complex phenomena through simple
objects [
Different scientists give different definitions of
an agent, but in general, an agent is a stand-alone
complex program with a detailed description of its
behavior, which is able to obtain information from
the environment and based on their experience to
respond correctly. Very often intellectual agents
are closely intertwined in the field of use with
artificial intelligence, but there is no complete
identity between them [
Sometimes, smart agents are considered in
combination with other agents. This structure is
called a multiagent system. It allows you to solve
a problem even more effectively. Agents need to
consider interacting with other agents for
cooperation or competition. That is, if a goal is
unattainable for one agent, then several agents can
cope with it, or if each of the agents is developed
to solve a specific problem, then the solution is the
one that should have the best result [
The most famous examples of intelligent agents in the world are Alexa (from Amazon), and Siri (from Apple). They process the user's request, collect data from the Internet and provide a response. They are often used to obtain information about the weather or weather forecast.
Professor Russell identified the following
main characteristics of the agent: survivability
(code works constantly and decides when to take
action), autonomy (the agent makes decisions
without human intervention), social behavior
(they can be involved through other components)
and reactivity (perceive and respond on the
context in which they are) [
There are two reasons that led to the
development of intelligent agents. The first is the
use of computer science. There are more and more
different technical devices in the world, such as
computers, servers, mobile phones, tablets, which
in turn can be connected to the network, and those
to the Internet. Previously, the number of
connection points was less than the number of
users, but today each of us can have several
different devices. Computing resources are
improving day by day, but the amount of data is
growing even faster. All this contributes to the
complexity of systems and their algorithms. To
facilitate data handling, systems are divided into
smaller subsystems. It is to solve such problems
that there are intelligent agents who study them at
a high level of abstraction. The second reason is
the development of society. Clever agents play a
significant role in analyzing patterns of human
interaction in different situations. People can
independently predict the behavior of other
people, conduct negotiations and discussions,
resolve conflicts, form organizational structures.
All this can also be analyzed and used by a smart
system. This is done by an intelligent agent who
can make decisions or execute assignments based
on experience, nested data and environment. They
can also be used to collect real-time information
[
The purpose of creating an agent model and
the process of building it depends on the order of
the decision maker. This order can be executed by
agents of several types. Today, intelligent agents
are divided into the following types [
Vicent J. presents three works related to
agentoriented programming [
The second paper proposes a new methodology for developing MAS work in semantic web environments. The proposed methodology is based on a specific area. A modeling language called the Agent Semantic Web Language. The training was demonstrated through a case study conducted using the wellknown JACK platform. The proposed example consists of a set of agents who exchange services or goods of the owners according to their preferences, without using any currency.
Finally, the third work presented the structure
of agent development for mobile devices. The
proposed structure allows users to create
intelligent agents with typical agent-oriented
attributes of social abilities, reactivity, proactivity
and autonomy. In fact, the main contribution is the
related support of the data framework. Supporting
related data corresponds to the ability to convey
the beliefs of the agent related data environment
and use these beliefs during the planning process
[
Previously, the effectiveness of the procedures
for reducing the compatibility of signals due to the
use of models of several reference forms on each
row of selection of a polynomial model was
proved [
In the process of using the described technologies to build a forecast model of the monitoring agent, it turned out that there are cases when the variety of proposed methods of model synthesis is not enough to adequately describe the processes of price changes on the stock exchange. Therefore, there is a need for additional research on the processes of synthesis of agent models with several reference forms for their adaptation to the conditions of a given subject area.
At the beginning of the synthesis of the model, the results of observations of the price of gold bonds at the close of trading on the stock exchange are known, where zt – the price of a gold bond at the time of the last observation, t – the value of the time of the last observation during the historical period;
which are recorded over a discrete period of time in one day: During historical period T:
Δt = t - t-1= t-1 - t-2 = 1.
T = {t, t-1, t-2,…, t-m}.
A predefined list of features of influencing factors that are used as independent variables.
X = {x1, x2, …, xn},
Zt+1 = f(X, T, Δt) where n – the number of signs of influencing factors.
It is necessary to build a forecast model (1) (2) (3) (4)
We propose to improve the method of model
synthesis using several reference forms in agent
synthesizers. The main task of the agent is to
transform information from a matrix of numerical
characteristics into the form of a model.
Depending on the simulation results, the system
issues a status change message. At the input, the
system adopts a multi-row GMDH algorithm and
a method of model synthesis, according to which,
with each row of selections, models with several
reference forms are generated and then the best
ones are selected. In the course of the research we
determined that, in contrast to the existing
method, where models of 6 reference forms were
generated on each row of selection [
Data for the array of observation results were
taken from financial exchange reports on the
Yahoo website [
As a simulated feature (dependent variable) used prices at the time of closing the exchange. The following were used as influential features (independent variables): ● the stock index, the basket of which includes the US joint stock companies with the largest capitalization, ● price of Australian dollar, ● exchange-traded investment fund specializing in treasury forms under fixed-term contracts of 7-10 years, The last 10 points formed a test sequence. These points did not participate in the creation of the model and were used to calculate the forecasting error.
In fig. 1 presents the test results of agent models.
Predicted observatNioenwpoint
Existing
The average signal error at the output of the agent model, built on the advanced method, became 0.9%. The error of the model built by the known method was 2.69%. Thus, the use in the synthesis process on each row of selection of models with reference forms, given in table. 1, allowed to reduce the average forecasting error by 10 points by 61.77%. market consolidated index of fixed income securities, ● US dollar index
In addition, we noticed that with a relatively small amount of data, multi-row GMDH can not accurately produce results, in contrast to the algorithm with models of two forms, which even with such a large amount of data could work at standard deviation 2.146009 (see Fig.2).
If you increase the amount of data to 500, the multi-row algorithm starts to work much better (standard deviation - 6.642833), but the proposed algorithm in this case works also good (standard deviation - 5.757732) (see Fig. 3).
150 100 50 0
Original
New
Standard 500 rows of data
Generation on each row of selection of models with several reference forms, allows to increase a variety of agent synthesizers. To adapt the processes of model synthesis according to the multi-row GMDH algorithm to the properties of the input data array, it is necessary to optimize the list of reference forms of models. For each input array, the list of reference forms of the models used in each row of selection must be determined separately.
Improving the process of building models with an agent synthesizer can increase the efficiency of the task of the agent as a whole. In addition, it was found that with a small amount of input data, an improved method of synthesis of models with two reference forms is able to build useful models with fewer observation points.
Future research will focus on the study of monitoring agents based on methods based on the combinatorial GMDH algorithm.