SELECTED

CONFERENCE PAPERS

Basic Trends of Decentralized Artificial Intelligence

V. I. Gorodetskya,*
a
Joint Stock Company EUREKA St. Petersburg, 196006 Russian Federation
*e-mail: vladim.gorodetsky@gmail.com

Abstract—Analysis of modern trends, advanced concepts, and projects in the field of information technolo-
gies shows the growing role of artificial intelligence, primarily decentralized artificial intelligence. The paper
analyzes the advanced concepts of building new generation applications in the field of information technol-
ogies and relevant applied developments and provides a brief analysis of modern achievements in the field of
decentralized artificial intelligence and self-organization, which are able to support the practical implemen-
tation of such applications. A simplified analog of the roadmap for the development of decentralized and self-
organizing artificial intelligence for the near future has also been formulated.

Keywords: IT applications, decentralized artificial intelligence, networked systems, self-organization, peer-

to-peer interactions, p2p-communication environment, decentralized machine learning, group control,
active data and knowledge
DOI: 10.1134/S105466182303015X

INTRODUCTION decentralized group control in real time. Military

applications, space, harmful environments, natural
A new wave of interest from the industrial com- disasters, and other areas are potential sources of
munity in the use of artificial intelligence (AI) tech- new IT applications and new requirements for AI
nologies has a strong impact on the priorities and capabilities.
trends of its development. The modern development
of the theoretical and technological basis of AI is What all these systems have in common, from a
mostly limited to centralized architectures and formal point of view is that they operate in a com-
stand-alone applications in the field of machine munication environment with a dynamic topology,
learning for computer vision, speech processing and and therefore each such application is a dynamic
voice assistants, natural language, classification, network.
and prediction tasks within business problems of
various purposes. AI applications of other classes are In the last few years, new projects and new para-
much less common and usually at the prototype digms of computing in dynamic networks have
level. There are many applications in the field of appeared in IT, which, on the one hand, rely on the
information technology (IT) that claim to use AI achievements of AI and, on the other hand, set new
without having convincing arguments for doing so. benchmarks and prospects for its development. As
However, the needs for IT in the use of the latest many authors believe, these landmarks are most con-
AI technologies are constantly growing, and this is centrated in the paradigm and architecture of peripheral
one of the notable trends in the IT industry, and the computing, in the concept of decentralized business
leader of these trends is decentralized AI [5]. This is without intermediaries with peer-to-peer (p2p) and
due to the features of a noticeable part of modern blockchain technologies at the core, in a new version
applications, which usually consist of a large number of the World Wide Web called Web3, and also in a proj-
of autonomous mobile entities that intensively inter- ect called the metaverse [5].
act in the process of collaborative problem solving.
These include a large number of new applications of These projects have already received some devel-
the Internet of things and group control of mobile opment [3, 5, 15, 24, 25, 34], and IT professionals
unmanned objects. The globalization of industry associate them with the future of IT technologies and
and business is also leading to the growing role of IT applications of the next-generation, and they also
distributed and decentralized architectures. There believe that the development of AI, at least in the
are and most likely will be new classes of relevant forthcoming years, will be determined by the require-
applications that will need to solve the problems of ments on their part. The value of these projects lies in
the fact that, on the one hand, they set in a concen-
trated form the basic requirements for the level of AI
Received January 30, 2023; revised January 30, 2023; technologies of the near future, and, on the other
accepted January 30, 2023 hand, they will become sources of new ideas and tech-

ISSN 1054-6618, Pattern Recognition and Image Analysis, 2023, Vol. 33, No. 3, pp. 324–333. © Pleiades Publishing, Ltd., 2023.
 BASIC TRENDS OF DECENTRALIZED ARTIFICIAL INTELLIGENCE 325

nologies and will stimulate the creation of new tools to systems can have a hybrid architecture. A detailed
support the development of a new generation of IT description of this architecture and its properties can
applications such as those which were listed above and be found in [5, 34].
even more difficult ones.
Decentralized business. In this concept, business
The paper [5] based on the analysis of a new chal- operates without intermediaries, i.e., on the basis of
lenges that are now being addressed by the IT para- direct peer-to-peer interactions of its participants. Its
digms focuses only on the importance of research and most developed version is the decentralized financial
development in the field of decentralized AI technol- business (DeFi). It is an alternative to the traditional
ogies. The objective of this paper is to take the next financial business and its instruments. In it, all trans-
step, namely, (1) to identify in these paradigms a basic actions with financial assets are carried out without
set of specific difficult problems, the solution of which intermediaries [5, 15].
requires the involvement of ideas and methods of
decentralized AI; (2) to conduct a brief analysis of the DeFi manipulates digital assets (DeFi tokens),
results, which decentralized AI has already at present; which are understood as entities, perhaps abstract,
and (3) to focus on some of the most important tasks, that have value in the financial market. One example
which it would makes sense to bring to the attention of of digital assets is cryptocurrencies. Another example
researchers and developers in the near future. In the
rest of the paper, Section 1 summarizes the objectives is nonfungible tokens (NFT1). They are similar to
and key features of the new projects and it paradigms securities, but they are digital. Not everyone can copy
mentioned above; Section 2 provides minimum infor- them, because they exist in their ecosystem.
mation about multi-agent systems (MASs) and self- The next concept of DeFi is a digital wallet, and in
organizing control principles that form the context fact, it is a user interface for accessing and managing
needed to understand the subsequent material. Sec- his assets. Another concept of DeFi is called a smart
tion 3 lists the achievements in the field of decentral- contract, which means the procedure for performing
ized AI and self-organization, which are now ready for
certain actions in the DeFi ecosystem, for example,
use. Section 4 analyzes promising research directions
in the field of decentralized AI and self-organization, the action to transfer a digital asset from one market
which will significantly expand the areas of their prac- participant to another. Manipulations with these DeFi
tical use in IT applications, at least in the near future, components are subject to certain rules, which must
ensuring their computational efficiency and fault tol- be automatically supported by the DeFi ecosystem.
erance. In Conclusion, the main results of the paper This support is done through the use of peer-to-peer
are summarized. interaction protocols. Examples are rules for perform-
ing actions such as “transaction generation,” “data
verification and decision making,” and “save the
1. NEW IT PARADIGMS AND data.” This DeFi chain of operations is performed
REQUIREMENTS FOR AI CAPABILITIES without any intermediaries [15].
Below is a summary of new projects, concepts, and The implementation of DeFi-business is based on
developments in the field of IT, which, according to blockchain technology [3]. Blockchain is a chain of
experts’ opinions, determine the requirements for the digital blocks following each other (transactions) con-
level of AI in the near future, at least [5]. nected by a common context. Note that the use of
blockchain technology is not limited to the storage of
Peripheral computing. This is the architecture of transactions. It is widely used in other tasks, which is
decentralized networked computing. In this architec-
due to a number of useful properties that this technol-
ture, data processing is carried out as close as possible
to their sources and to the users of the results. This ogy has. These include decentralized storage and exe-
architecture is of particular importance for mobile cution of operations with digital assets, transparency
networks, since the transfer of computing to the (access to details and to the trajectory of transactions),
periphery increases the speed of data processing and immutability of data during storage, the absence of
reduces delays caused by the limited resources of intermediaries (their role is played by consensus pro-
mobile devices, including power consumption [34]. tocols), anonymity (authentication of the user without
Systems built on this architecture can generally have revealing his identity), the ability to audit, built-in
different combinations of centralization and decen- protection against attacks, and fault tolerance.
tralization. Individual local subsystems can use a 1 An
decentralized architecture, but if there is a hierarchy in example of NFT is digital art. Beeple’s digital work,
5000 days, was sold at auction for 69 million dollars; see
the network, some functions can be performed in a https://www.heverge.com/2021/3/11/22325054/beeple-christies-
centralized version. Therefore, peripheral computing nft-sale-cost-everydays-69-million.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

326 GORODETSKY

Summarizing the above description of the features virtual world and their connections with similar com-
of decentralized business, we can conclude that the ponents of the physical world, thus gradually forming
support of decentralization based on peer-to-peer a virtual world and a single physical-virtual world with
interactions of objects in a p2p-communication envi- the metaverse as its part.
ronment determines the essence of the requirements The worlds interact with data. Data of the physical
put forward by them in relation to AI. environment from sensors, from social networks,
The concept of Web3 [24]. This concept aims to media, etc., are fed to the input of digital twins, chang-
develop the World Wide Web toward decentralizing ing their states and initiating dynamic processes. This
content storage. Due to this concept, IT giants such as data can be used for a variety of purposes, including
Google, Amazon, and Apple will lose their monopoly machine learning. Digital twins can generate model
on the storage of content and control over it. The data about the real world, which can replenish infor-
owner of the content in Web3 stores it on his own side, mation about the real world, creating a more complete
and therefore this content is protected from external picture of it in this way. This allows us to more accu-
blocking. This increases the level of trust in content rately solve the problems of predicting events, facts,
and its security, makes less vulnerable the transfer of and processes of the real world.
digital businesses such as DeFi to the Internet envi- Analyzing the requirements for AI technologies
ronment, and expands the opportunities of the digital necessary to create and use the metaverse, we can con-
economy based on tokens as digital assets. It is com- clude that they combine the requirements of other
monly said that Web3 is a transition from human texts concepts from the list above, and even these concepts
to texts for computers that can analyze them, make themselves.
conclusions, use them in the interests of machine
learning, etc. Obviously, in the concept of Web3, the
term decentralization is also key in terms of the 2. MULTI-AGENT TECHNOLOGIES
required properties of IT and AI technologies. AND SELF-ORGANIZATION:
Metaverse. This project expands the concept of GENERAL INFORMATION
cyberspace, adding to it digital twins of the compo- The analysis of the requirements for the directions
nents and aspects of the physical world and the inter- of development of AI from the side of next generation
action between them. This leads to the creation of a of IT applications, the composition of which can be
virtual world parallel to the physical world [25]. Judg- imagined according to the content of the previous
ing by the stated goals, the project is preparing a digital section, is advisable to carry out, supported by at
explosion of cyberspace. Its predecessors are virtual least minimal assumptions about the properties of
environments such as social networks, video confer- these applications, for example, about architecture.
ences, augmented reality systems, etc. As noted in the introduction, the new most popular
In the metaverse, the people of the physical world and important classes of IT applications are usually
interact with each other and with this environment network structures of dynamic topology with a large
through their intermediaries—digital avatars [25]. The number of nodes and links of various semantics.
creation and development of the application of the These networks are characterized by the fact that
metaverse takes place in three stages: their different nodes possess different information
and are able to solve different problems; they inten-
− creating digital twins in a virtual environment; sively interact with each other and the external envi-
− the settlement of its “digital aborigines” and the ronment in solving various problems, many of which
development of digital twins with their participation; are solved jointly by groups of nodes. In the architec-
tures of their software implementation, network
− parallel coexistence, interaction, and mutual nodes are usually put in line with the objects of the
enrichment of both worlds and their practical use [25]. application, for example, unmanned aerial vehicles,
According to this scheme, some set of digital twins robots in collective robotics, satellites in the space-
are first created for a number of aspects of the real based surveillance system, etc.
world. At this stage, they are called “shadows” of For such systems, the most suitable architecture
aspects of reality in the virtual world. Further, real- and development technology is the architecture and
world humans, through their avatar digital aborigines, technology of the autonomous agent and MAS. This is
create connected ecosystems, encompassing such due to the fact that the concept of an agent on default
aspects of the physical world as science, culture, eco- assumes autonomy and the ability for proactive behav-
nomics, laws, and social norms. They are created by ior as its basic properties. Autonomy and proactivity
analogy with the same ecosystems of the physical together make it possible to formalize the complex
world. At the beginning of the development of the behavior of many autonomous, possibly mobile net-
metaverse, the digital twins are loosely coupled with work objects, which depends not only on the current
similar physical worlds, but over time, digital aborigi- inputs of each object but also on the prehistory of the
nes establish connections within the ecosystems of the state of the external environment and on the behavior

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

 BASIC TRENDS OF DECENTRALIZED ARTIFICIAL INTELLIGENCE 327

of other objects of the system and their neighbors on tivity to initial states and small variations of parame-
the network. For example, a proactive agent can gen- ters with a jump change of state, multiplicity of stable
erate messages even if there are no input events, e.g. states (attractors), adaptability as the ability to
due to a timeout exceeding. Another important prop- change behavior and structure when changing local
erty of the agent, essential for the implementation of inputs, and the potential complexity of the behavior
networked systems comprising many objects, is its of simple network objects as a consequence of their
interactivity. Interactivity is defined as the ability of large number.
agents to exert particular influence on each other, and
it is in this sense that the network of MAS agents is Most of the developed prototypes of self-organiz-
called “weakly-coupled.” As a side effect of agents’ ing systems are implemented in the MAS architecture,
autonomy and interactivity, their ability to solve com- and this is not accidental. Indeed, the basic require-
plex problems throgh cooperative coordinated behav- ments for the software implementation of such a sys-
ior emerges [12]. tem are that its components must be autonomous (be
From this brief listing of agent properties, it follows able to control their own behavior aimed at achieving
that the agent model for network node individual their local goals without external intervention). They
behavior specification and the MAS model for speci- should be able to perceive the external world and
fication of networked object behavior as a whole are locally effect it, be able to interact with their neighbors
generally well suited for IT applications of the class in (on the network, in space, etc.), maintain the emer-
question. In addition, another convincing argument in gent structure of the system, and have the means to
favor of such a choice is that the behavior of the proac- control their behavior. Until now, the MAS has been
tive agent is usually specified by the finite-state and remains the alternativeless architecture for the
machine, and the network of such interacting agents is software implementation of self-organizing systems.
conveniently modeled by a network of such interacting Moreover, at present, the development of principles
finite-state machines. and models of self-organization takes place within the
In the described architecture, a software agent rep- framework of research and development in the field of
resenting a particular stand-alone application object, MAS [11, 17].
on the one hand, controls the internal behavior of the
software and/or hardware components of “its” own
network node in various use cases. On the other hand, 3. DECENTRALIZATION
this software agent is a representative of “its node” in AND SELF-ORGANIZATION OF AI SYSTEMS.
a network of software agents, where its function is to LEVEL OF MODERN ACHIEVEMENTS
interact with agents of other nodes of the network by
exchanging messages to coordinate the behavior of “its In studies on the designated topic, two waves of
own node” in the joint solution of some common task. activity can be distinguished. The first of these covers
Large-scale network structure systems with a the period from about the mid-1980-ies to the end of
decentralized storage and computing architecture typ- the 1990s. During this period, the main themes of
ically allow only the interaction of “neighboring” these research and active development, which were
nodes of the network (agents representing these nodes funded by DARPA (USA), were the problems of group
in the network software environment), for example, in control of teams of autonomous agents performing a
resource planning and group control, regardless of common mission. The second wave dates back to
whether the “neighborhood” is determined by the 2002–2010, and its main focus was on two key prob-
structure of physical communication channels lems of decentralized AI, namely, decentralized
between them or, for example, overlay network in the machine learning and the problems of creating p2p-
case of a software-defined network. The individual infrastructures to support the p2p-interaction of
node of the network (the agent representing it) may autonomous network agents in a p2p-communication
not know at all about nodes (agents) that are not its environment. Around 2000, active research and devel-
neighbors. Therefore, all calculations in such networks opment began in the field of self-organization as the
are performed on the basis of local interactions and basic principle of adaptive control in large-scale appli-
self-organization principles. cations of the network structure. Active research in
Self-organization is defined as a dynamic process of this area continued until about 2015. It can be argued
a system, implemented in it without external interfer- that four of these problems (group control, decentral-
ence on the basis of local interactions of its objects, ized machine learning, p2p-communication net-
which leads to the emergence and maintenance of a works, and self-organization) at different times were
structure on a set of its objects [17, 33]. Self-organiz- the focus of research related to decentralized AI. Other
ing systems also have a number of specific properties, relevant problems are more specific and the results
which include autonomy, global order arising owing to obtained in them are not analyzed here. Let us take a
local interactions, emergent behavior, possible insta- look at the main results so far on the four issues men-
bility that occurs without external inf luences, sensi- tioned.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

328 GORODETSKY

In the field of group control (teamwork), by the their significant multiyear financial support from
mid-1990s, several theories were proposed, but only DARPA and unambiguous focus on military applica-
two of them laid the theoretical basis for subsequent tions, both developments were closed by the early
and modern models and software tools in this field. 2000s because the models proposed in them proved
Pioneering was the work [9], in which the Joint Inten- unsuitable for practice due to enormous computa-
tions Theory was proposed. It formulates the basic tional complexity, since both were specified based on
concepts and general framework that define the group predicate calculus, extended by modal and temporal
behavior of agents and the characteristics of their operators.
interaction in this behavior, as well as the principles of
information exchange that can support their situa- For the subsequent period until about 2010, there
tional awareness necessary for decentralized coordi- was a certain stagnation in the development of the the-
nation of their individual behavior in order to achieve ory and models of group control and its applications,
the group goal. Apparently, the most important and which is clearly noted in [14]. During this period,
practically very useful result of this theory is the proto- researchers tried to adapt the theoretical models
col of interaction of members of the team of agents, described above to practical needs. Among the fortu-
called the joint intentions protocol [22]. It is used by nate models of this period is the BITE model pro-
team agents to agree on their commitments and group posed in [23]. In it, the individual behavior of team
conventions and to solve common issues of distributed agents and group control are modeled by three struc-
coordination of their behavior in an autonomous mis- tures. The first is the hierarchical structure of the tasks
sion without outside intervention. of the team of agents (a structured plan of their
Another theory, known as the Shared Plans Theory, actions). The second describes the structure of the
is constructed a little differently [21]. Its basic con- agents and their subgroups, which are assigned to the
cepts are the group plan and the individual mental individual mission tasks described in the first struc-
concepts of autonomous agents. This group plan, in ture. The third structure explicitly describes the com-
addition to the set of actions of individual agents of the munications and interactions of agents in the process
group, which agreed with the set of conditions (time, of distributed coordination of agent behavior. This
place, resources, etc.), contains infrastructure compo- model significantly simplifies the classical models [9,
nents of the model, which convert the set of distrib- 21], in which the scenario model is not specified
uted agents into a single team. Both theories have a explicitly before the start of the execution of the team’s
strict mathematical justification, but so far their soft- mission, but must be output dynamically. But, never-
ware implementations are limited to software develop- theless, the BITE model also has a big drawback. The
ment at the level of simple prototypes. authors note that the essence of the group control
automation process is to provide automatic control of
In subsequent developments, different authors distributed scenario execution according to some stan-
used different combinations of individual ideas of both dard protocol. They note that this goal in the BITE
theories for their models. The two most well-known model has not been achieved, although even more is
are STEM [31] and RETSINA [29]. To support the required to ensure self-organization of group control.
software implementation of the STEM model, the This issue has been resolved in a later work [14].
Teamcore environment was developed [30, 31]. In it,
the architecture of the software agent that manages the The period from 2002 to 2020 is characterized by
behavior of the agent is divided into two parts. One of active research in the field of decentralized AI in clus-
them is subject domain-dependent, and the other is ter analysis tasks and in decentralized machine learn-
subject domain-independent. The domain-indepen- ing tasks, although in fairness it is worth noting that in
dent part of the agent is called a teamcore agent. It plays the United States decentralized machine learning
the role of “wrapper” for its domain-dependent part. based on a set of distributed data was started in the
The wrapper is responsible for the external behavior of early 1990s. For example, interesting theoretical
the agent and provides it with the capability to work in results brought to practical use were obtained in the
teams. Note that this productive idea was later devel- works [8, 28]. The authors of these works already then
oped in different works. solved the problem of decentralized learning to detect
The RETSINA model [29] has an architecture that false transactions in a group of US banks, which did
exploits mainly the Shared Plan Theory. In its soft- not agree to provide their data to machine learning
ware architecture, an agent is allocated, which is called specialists, but agreed that training would be con-
a cooperative interface agent. It actually implements the ducted by distributed team of agents using local data of
centralization of group control. In an autonomous each bank separately (in a decentralized version) with
mission, this solution can be critical, for example, the subsequent use of a common set of rules for detect-
when an object on which an agent with the interface ing false transactions based on local data and on meta-
agent role is installed fails. Both models and their sup- data. The method that was developed by the authors of
porting toolkits have long been regarded as world lead- these works was successful and was used already then
ers in group control theory and models. But despite in practice.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

 BASIC TRENDS OF DECENTRALIZED ARTIFICIAL INTELLIGENCE 329

In the period (2002–2006), an innovative project new properties of next generation applications have
of the European FP6-IST program was carried out been significantly tightened, in particular, with regard
under the name “KD-Ubiq: A Blueprint for Ubiquitous to their scalability, fault tolerance, and computational
Knowledge Discovery Systems.” This project explored complexity, as well as trust in the results obtained.
the difficult problem of decentralized data mining and
machine learning algorithms with an emphasis on pre-
serving the confidentiality and privacy of data. The 4. TOPICAL PROBLEMS
results of this project were published quite fully in the OF DECENTRALIZED AI
book [26]. Some results, which still retain their great It should be noted that the main results in the field
relevance and practical significance, are published in of decentralized AI, despite the sufficient theoretical
[10, 32]. In particular, these works proposed protocols maturity, currently cannot be directly used in applica-
for decentralized calculation of mathematical expec- tions of the new generation, since the latter impose
tation, as well as an approximate algorithm for decen- different requirements on the development than those
tralized clustering, built on the basis of the K-means that were characteristic of applications ten years ago.
clustering method. In both protocols, each agent First of all, these requirements relate to the scale of
receives meta-information only from its network applications—modern and promising applications are
neighbors. much larger in scale than those that were focused on
It is also worth paying attention to the active devel- the development period (1990–2020). This applies to
opment of decentralized machine learning methods all four issues discussed in the previous section. As a
and models at that time, which were focused on the result of the large scale of modern applications, prob-
use of agents and MASs technology of decentralized lems associated with ensuring application computa-
machine learning. This direction was called Agent tional efficiency, robustness, security, and a number of
Mining [2, 7]. In [6, 27], there is a review that describes others mentioned below are exacerbated.
the problems of this direction, as well as the most A noticeable increase in the scale of applications
interesting decentralized machine learning algorithms leads, first of all, to a worsening the robustness of cal-
developed by that time. culations performed in a decentralized architecture.
This problem is not yet given its due, although it is a
Work on creating a p2p-platform to maintain the critical issue in conventional distributed computing, in
interaction and communication of autonomous dis- the computing of statistics, and in decentralized
tributed entities (agents) without centralized yellow machine learning algorithms. In a new way, the scale
pages was initiated in 2004 by the FIPA working of systems affects the use of supercomputers. For
group2 [13], and already in 2007, the first software example, the well-known parallel computing technol-
implementation of such a platform was published in ogies implemented in the Hadoop ecosystem are not
accordance with the reference model of FIPA [20]. On suitable for use in the supercomputers, which are able
its basis, during this period, several fairly indicative to realize their capabilities only if all the data is in
fully decentralized AI applications were developed RAM, when there is almost no need to exchange data
([17, 19], etc.). with external memory. For these reasons, many
decentralized computing algorithms in next-genera-
As for self-organization, the fourth important tion applications require either modification or com-
direction of research in the field of decentralized AI, plete renovation.
the current state of research in this field is covered in
detail in the works [17, 18, 33]. This direction has Decentralization also brings its own problems,
been studied quite deeply and there are many practi- among which the security of distributed components
cal developments in this area in the period until and data trasmission channels is a priority. New tech-
about 2010. Currently, this topic is not so popular in nologies always bring new problems and the problem
theoretical studies. However, there is still a great field of security of decentralized systems is one of such new
and a great potential for its development in this direc- problems.
tion, and some aspects of this type are discussed in New challenges for IT applications include the
the next section. development of new infrastructure and tools. For now,
in the current implementation of IT applications, the
Thus, it can be argued that to date, the results share of decentralized AI is very small. And the rea-
obtained in the field of decentralized AI relatively fully sons for this lie, first of all, in the fact that decentral-
cover its main scientific areas and are quite mature. ized AI technologies are not yet ready for use at the
However, these results are mainly from the period up industrial level. The list of priority developments in
to 2010 and they meet the requirements of the applica- this area is as follows.
tions of that time. By now, classes of IT applications in
which AI technologies could lead to fundamentally 1. Development of new robust algorithms, mature
technologies, and scalable software tools for the
2 FIPA—Foundation
for Intelligent Physical Agents, implementation of the concept of p2p-interactions of
http://www.fipa.org/. autonomous objects in networks with dynamic con-

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

330 GORODETSKY

nectivity and intensive message passing with large vol- oretical proposals, as well as specific developments,
umes of transmitted information. These tools must that allow for the construction of self-organizing sys-
provide reliable communications with the temporary tems of decentralized architecture with the involve-
loss of availability of individual addressees without loss ment of a wider and even global context. Examples are
of addressable information. The creation of new self-organization systems using digital fields [4], as
industrial-level platforms to support p2p-communi- well as self-organization models using the concept of
cations and dynamic routing in large-scale networks amorphous computing [1].
and variable topology is one of the important tasks of
decentralized AI in the near future. These and other approaches for the formation of a
global context based on local information are united
2. Development of known and creation of new by the concept of active data and knowledge bases. By
robust algorithms, technologies, and software tools to definition, a data and knowledge base is called an
support decentralized processes of data mining and active database if it can perform not only the actions
machine learning operating, in partiacular, in real that the user explicitly specifies. They can perform
time. In mobile networks with limited computing and other actions of a proactive nature in accordance with
communication resources, this will significantly the rules (knowledge) embedded in the data model.
reduce the burden on communication channels and Typically, the activity of the data and knowledge
accelerate learning processes. To solve the described model is used to control and maintain its consistency
problems, there is already a necessary theoretical and integrity. In some cases, the active knowledge
basis, experimental software developments have been model includes the calculation of some attributes. If
created, but computationally efficient robust indus- the values of these attributes fall within a predeter-
trial-level tools for this purpose have yet to be created. mined range, this fact triggers proactive behavior of
3. Creation of scalable algorithms and technologies the system. A typical example of such proactivity in a
to support applied p2p-services, as well as ecosystems distributed system is a timeout, which controls the
of such services, including, for example, decentralized correctness of the system processes in time and gener-
planning services, p2p-services of distributed coordi- ates certain control actions in case of violation of
nation of group behavior of objects solving a common thresholds specified by timeouts.
problem, and other services of an applied nature.
The concept of active knowledge is very useful in
4. A complicated and important complex of prob- self-organizing systems. Expanding the locally avail-
lems of decentralized AI will have to be solved in rela- able context about the states of network objects to
tion to the tasks of group control. This is a new field of make decision based on active knowledge is one of the
research and development, in which the theoretical options for increasing the autonomy and situational
basis is based on the protocols, i.e., decentralized awareness of distributed decentralized systems. We
(p2p-) algorithms of external behavior of groups of demonstrate the ability of active knowledge to expand
agents in various use cases. A decentralized algorithm the local context of decision-making mechanisms of
is a protocol of interaction of distributed objects, self-organization by example.
which allows them to coordinate their individual
behavior within the framework of the scenario of Example: Dynamic routing in a network with
group behavior of agents intended to solve a common dynamic topology. We consider the concept of amor-
problem. Essentially, these are consensus protocols, phous computing proposed at MIT in 2000 [1]. It is
leader selection protocols, contract network protocol, based on a model of self-organization borrowed from
auction protocols, a protocol of joint intention, proto- morphogenesis. This model considers a vector model
cols of information exchange in order to maintain sit- of the morphogen, in which each coordinate can be
uational awareness of teammates in group behavior, used to control a particular process, property, etc.
and other protocols of decentralized computing, such The amorphous computing model considers a
as decentralized algorithms for data mining and large number of equally programmed simple devices
machine learning—all of them can and should become that are distributed on the surface or in some volume
components of libraries of standard algorithms for randomly. Each device can perceive the external envi-
decentralized computing and group control of net- ronment and effect on it. It is believed that the devices
works of autonomous objects represented in the soft- have very limited resources and perceive a very limited
ware environment by their agents. amount of local information, and they can also fail. It
5. Algorithms and technologies of self-organiza- is also assumed that each device has its own execution
tion that work in a broader context, up to the global thread and is capable of generating random numbers.
one. Decentralized AI algorithms form the basis of Each device has an internal state depending on its pre-
self-organization. However, the modern concept of vious actions. Devices can exchange messages over a
self-organizing algorithms involves the use of only communication channel with a short range of reach.
local information, i.e., information that an agent can Network devices initially know nothing about the
obtain from its neighbors, and its use in local optimi- topology of the communication network; the network
zation processes. However, there are a number of the- does not have a centralized source of information,

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

 BASIC TRENDS OF DECENTRALIZED ARTIFICIAL INTELLIGENCE 331

global time, and beacons for binding to coordinates. New challenges for IT applications, as well as new
The communication environment supports the propa- requirements and new opportunities for decentralized
gation of the digital field of each device, which is AI, are a natural development process. At this stage,
defined by a spatially dependent data structure called the issues and tasks summarized in this section are
a digital field [4]. considered as the key ones.
It turns out that such a fairly simple computing
model can be used to build very effective mechanisms CONCLUSIONS
of self-organization by spreading part of the global
context across the network. For example, let some The large scale, high level of complexity, and dis-
source device can send its name and “morphogen” to tributed nature of the new generation of IT applica-
neighbors—a number equal to zero. After receiving tions and the tasks that they must solve, as a rule, in
such a message, each neighbor sends it to its neigh- real time and in uncertainty conditions is a modern
bors, adding 1 to the value of the morphogen with the reality. Complex next-generation IT applications are
name of its source. This process continues until all networks of mobile objects operating in a wireless
nodes in the communication network are reached. communications environment. These objects, owing
to the combination of their resources and intensive
Each device can store the minimum value of this interaction, are able to solve the most complex prob-
morphogen. If necessary, the device can use it as the lems even with limited computing resources. For
value of the shortest path to the source node, if the modern AI, this situation is new, and it presents new
path length is measured in terms of hops on it. Further, requirements that cannot be addressed without revis-
each node can determine the “local orientation” of the ing previously accepted control paradigms and
node, i.e., the direction to or from the source, request- approaches and without tightening various indicators
ing the value of the morphogen of this node from of the quality of the system.
neighbors. The direction to the source is determined Analysis of modern trends in the field of advanced
by the node (or nodes) in which the target node mor- concepts and projects under development shows the
phogen value is minimal. Similarly, the direction from growing role of AI, primarily decentralized AI.
the source corresponds to the direction of nodes with
a value of the morphogen one more than its own value. The paper analyzes the advanced concepts of
building new generation IT applications and specific
If network agents are used as digital field sources, developments and provides a brief analysis of modern
the agent of any network node can obtain information achievements in the field of decentralized AI and self-
about the “local orientation” of the network node to organization, which are theoretically able to support
the source by polling neighboring nodes. If the source the practical implementation of such applications.
is the addressee of the message that a node must send, However, these developments are not yet ready for use
then as a result of this survey it can dynamically deter- at the industrial level, taking into account new realities
mine the first hop of the route of transmission of the and requirements for them from modern applications.
message to its addressee. Next, subsequent nodes on The basic guidelines for the development of AI in this
the path from source to destination can do this. As a context can be characterized by such key areas of
development as the following:
result, the desired route of transmission of the message
will be formed automatically in a decentralized style − decentralized computing in dynamic networks
using an amorphous computing model. based on algorithms of p2p-interactions;
This example shows the important role of active − dynamic p2p-communication networks and
knowledge in solving the problem of addressing mes- environments;
sages in systems with mobile objects and limited com- − group behavior modeling and protocol-based
munication range. Note that, in this case, the local group control;
context of the network node is expanded owing to a
digital field that delivers distributed information about − libraries of application protocols for interaction
the global connectivity of the communication network of agents in different tasks and in cases of use and eco-
to each node. systems that support the operation of objects in accor-
dance with these protocols;
Thus, the use of active knowledge in distributed
systems with self-organization allows one to get more − large-scale robust computationally efficient
effective self-organizing systems by “delivering” infor- algorithms for decentralized computing in p2p-net-
mation about the global context to local decision- works;
making nodes. Therefore, the use of active knowledge − new decentralized algorithms, technologies, and
should be considered as one of the important future software tools to support data mining and machine
principles of building decentralized AI systems. learning processes;

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

332 GORODETSKY

− scalable algorithms and technologies to support 11. G. Di Marzo Serugendo, M. P. Gleizes, and A. Kara-
applied p2p-services, as well as ecosystems of such ser- georgos, “Self-organisation in multi-agent systems,” in
vices; Rapport de recherche IRIT/2005_18_R (Univ. Paul Sa-
batier, Toulouse, 2005).
− self-organization in a global context.
12. J. Ferber, “Foreword,” in Handbook on Agent-Oriented
Design Processes, Ed. by M. Cossentino, V. Hilaire,
A. Molesini, and V. Seidita (Springer, Berlin, 2014).
CONFLICT OF INTEREST https://doi.org/10.1007/978-3-642-39975-6
The author declares that he has no conflicts of interest. 13. FIPA P2P NA WG6. Functional Architecture Specifica-
tion Draft 0.12. http://www.fipa.org/subgroups/P2PNA-
WG-docs/P2PNA-Spec-Draft0.12.doc.
REFERENCES 14. K. Geihs, “Engineering challenges ahead for robot
teamwork in dynamic environments,” Appl. Sci. 10,
1. H. Abelson, D. Allen, D. Coore, C. Hanson, G. Hom- 1368 (2020).
sy, T. Knight, R. Nagpal, E. Rauch, G. Sussman, and https://doi.org/10.3390/app10041368
R. Weiss, “Amorphous computing,” Commun. ACM
43 (5), 74–82 (2000). 15. I. Gidaspov and A. Akopyan, “What is DeFi: Every-
https://doi.org/10.1145/332833.332842 thing about decentralized finance.” https://curren-
cy.com/ru/chto-takoe-defi
2. Agent Mining Special Interest Group.
http://www.agentmining.org. 16. V. Gorodetsky, “Multi-agent autonomous group con-
trol in collective robotics-based assembly,” CEUR
3. M. N. M. Bhutta, A. A. Khwaja, A. Nadeem, H. F. Ah- Workshop Proc. 2648, 1 (2020).
mad, M. Kh. Khan, M. A. Hanif, H. Song,
M. Alshamari, and Yu. Cao, “Survey on blockchain 17. V. I. Gorodetskii, “Self-organization and multiagent
technology: Evolution, architecture and security,” systems: I. Models of multiagent self-organization,”
IEEE Access 10, 61048–61073 (2021). J. Comput. Syst. Sci. Int. 51, 256–281 (2012).
https://doi.org/10.1109/ACCESS.2021.3072849 https://doi.org/10.1134/s106423071201008x
18. V. I. Gorodetskii, “Self-organization and multiagent
4. M. Camurri, M. Mamei, and F. Zambonelli, “Urban systems: II. Applications and the development technol-
traffic control with co-fields,” in Environments for Multi- ogy,” J. Comput. Syst. Sci. Int. 51, 391–409 (2012).
agent Systems III, Ed. by D. Weyns, H. V. D. Parunak, https://doi.org/10.1134/s1064230712020062
and F. Michel, Lecture Notes in Artificial Intelligence,
Vol. 4389 (Springer, Berlin, 2007), pp. 239–253. 19. V. I. Gorodetskii, O. V. Karsaev, V. V. Samoilov, and
https://doi.org/10.1007/978-3-540-71103-2_14 D. Trotskii, “Development tools for open agent net-
works,” J. Comput. Syst. Sci. Int. 47, 429–446 (2008).
5. L. Cao, “Decentralized AI: Edge intelligence and smart https://doi.org/10.1134/S1064230708030131
blockchain, metaverse, Web3, and DeSci,” IEEE In-
tell. Syst. 37 (3), 6–19 (2022). 20. V. Gorodetsky, O. Karsaev, V. Samoylov, and S. Sere-
https://doi.org/10.1109/mis.2022.3181504 bryakov, “P2P agent platform: Implementation and
testing,” in Agents and Peer-to-Peer Computing, Ed. by
6. L. Cao, V. Gorodetsky, and P. A. Mitkas, “Agent min- S. R. H. Joseph, Z. Despotovic, G. Moro, and S. Ber-
ing: The synergy of agents and data mining,” IEEE In- gamaschi, Lecture Notes in Computer Science,
tell. Syst. 24 (3) (2009). Vol. 5319 (Springer, Berlin, 2010), pp. 41–54.
https://doi.org/10.1109/MIS.2009.45 https://doi.org/10.1007/978-3-642-11368-0_4
7. L. Cao, G. Weiss, and P. S. Yu, “A brief introduction to 21. B. J. Grosz and S. Kraus, “Collaborative plans for com-
agent mining,” Auton. Agents Multi-Agent Syst. 25, plex group action,” Artif. Intell. 86, 269–357 (1996).
419–424 (2012). https://doi.org/10.1016/0004-3702(95)00103-4
https://doi.org/10.1007/s10458-011-9191-4
22. N. R. Jennings, “Controlling cooperative problem
8. Ph. K. Chan, W. Fan, and S. J. Stolfo, “Distributed solving in industrial multi-agent systems using joint in-
data mining in credit card fraud detection,” IEEE In- tentions,” Artif. Intell. 75, 195–240 (1995).
tell. Syst. 14 (6), 67–74 (1999). https://doi.org/10.1016/0004-3702(94)00020-2
https://doi.org/10.1109/5254.809570
23. G. A. Kaminka and I. Frenkel, “Towards flexible team-
9. P. Cohen and H. J. Levesque, “Teamwork,” Noûs, work in behavior-based robots,” in Proc. 4th Int. Joint
No. 25, 487 (1991). Conf. on Autonomous Agents and Multiagent Systems (As-
https://doi.org/10.2307/2216075 sociation for Computing Machinery, New York, 2005),
pp. 1355–1356.
10. S. Datta, K. Bhaduri, C. Giannella, R. Wolff, and https://doi.org/10.1145/1082473.1082770
H. Kargupta, “Distributed data mining in peer-to-peer
networks,” IEEE Internet Comput. 10 (4), 18–26 24. A. Karelina, “What is Web3: Explanation in simple
(2006). words.” https://secretmag.ru/enciklopediya/chto-ta-
https://doi.org/10.1109/mic.2006.74 koe-web3-obyasnyaem-prostymi-slovami.htm

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023

 BASIC TRENDS OF DECENTRALIZED ARTIFICIAL INTELLIGENCE 333

25. L. Lee, T. Braud, P. Zhou, L. Wang, D. Xu, Z. Lin, Vladimir Ivanovich Gorodetsky.
A. Kumar, C. Bermejo, and P. Hui arXiv:2110.05352 Doctor of Technical Sciences
(2021). (1973), Professor (1990), Honored
Scientist of the Russian Federation.
26. M. May and L. Saitta, Ubiquitous Knowledge Discovery: He graduated from the Leningrad
Challenges, Techniques, Applications, Lecture Notes in Air Force Engineering Academy
Computer Science, Vol. 6202 (Springer, Berlin, 2010). with the master degree in mechanics
https://doi.org/10.1007/978-3-642-16392-0 (1960) and the Faculty of Mathe-
matics and Mechanics of the Lenin-
27. Ch. Moemeng, V. Gorodetsky, Z. Zuo, Yo. Yang, and grad State University with the mas-
Ch. Zhang, “Agent-based distributed data mining: ter degree in mathematics (1970).
A survey,” in Data Mining and Multi-agent Integration, He defended his candidate (1967)
Ed. by L. Cao (Springer, Boston, 2009), pp. 47–58. and doctoral (1973) dissertations on
https://doi.org/10.1007/978-1-4419-0522-2_3 the problems of optimal control. From 1967 to 1988 he
worked at the Mozhaisky Military Space Engineering
28. A. L. Prodromidis, Ph. K. Chan, and S. J. Stolfo, “Me- Academy in research and teaching positions. From 1988 to
ta-learning in distributed data mining systems: Issues August 10, 2018, he worked at the St. Petersburg Institute
and approaches,” in Advances of Distributed Data Min- for Informatics and Automation of the Russian Academy of
ing (AAAI Press, 2000). Sciences as head of the laboratory and chief researcher of
the laboratory of intelligent systems. He published more
29. K. Sycara, M. Paolucci, M. Van Velsen, and J. Giam- than 250 works (over 100 in foreign cited publications) and
papa, “The RETSINA multi-agent infrastructure,” 9 monographs and manuals; he was coeditor of 12 issues of
Auton. Agents Multi-Agent Syst. 7, 29–48 (2003). the “Lecture Notes in Artificial Intelligence” and “Lecture
https://doi.org/10.1023/a:1024172719965 Notes in Computer Science” series (2001–2015), guest edi-
30. M. Tambe, W.-M. Shen, M. Mataric, D. Pynadath, tor of a special issue of the IEEE Intelligent Systems journal
D. Goldberg, P. Modi, Z. Qiu, and B. Salemi, “Team- (Q1) on “Agent and Data Mining Interaction” (2009), etc.
work in cyberspace: Using TEAMCORE to make Member of the Russian and European Association of Arti-
agents team-ready,” in AAAI Spring Symp. on Agents in ficial Intelligence, IEEE, IEEE Computer Science, Inter-
Cyberspace (1999), pp. 136–141. national Foundation for Autonomous Agents and Multi-
agent Systems (IFAAMAS), International Society of Infor-
31. M. Tambe, “Towards flexible teamwork,” J. Artif. In- mation Fusion (ISIF). Associate Editor of the Data Science
tell. Res. 7, 83–124 (1997). and Analytics international journal (Springer) and a mem-
https://doi.org/10.1613/jair.433 ber of the editorial board of the Design Ontology Russian
journal. Winner of the D.A. Pospelov Award of the Russian
32. R. Wolff and A. Schuster, “Association rule mining in Association of Artificial Intelligence (2022).
peer-to-peer systems,” IEEE Trans. Syst., Man Cy- Science metric data (as of January 1, 2023): RSCI cita-
bern., Part B (Cybern.) 34, 2426–2438 (2004). tion 3166, h-index = 24, WoS citation 219, h-index = 7,
https://doi.org/10.1109/tsmcb.2004.836888 Scopus-citation 755, h-index =12.
Areas of scientific interest: theory of optimal control,
33. D. Ye, M. Zhang, and A. V. Vasilakos, “A survey of self- celestial mechanics, applied statistics, planning and schedul-
organization mechanisms in multiagent systems,” ing, artificial intelligence and decision-making, multi-agent
IEEE Trans. Syst., Man, Cybern.: Syst. 47, 441–461 systems, group management, self-organizing agent networks,
(2017). software tools, data mining and machine learning, efficient
https://doi.org/10.1109/tsmc.2015.2504350 and robust big data processing algorithms, semantic comput-
ing, distributed and p2p-learning, computer security, dealing
34. Ya. Zhang, “Mobile edge computing,” in Mobile Edge with uncertainty, recommendation systems, transport and
Computing, Simula SpringerBriefs on Computing, production logistics, steganography, etc.
Vol. 9 (Springer, Cham, 2021), pp. 9–21. Web page: https://en.wikipedia.org/wiki/Vladimir_Goro-
https://doi.org/10.1007/978-3-030-83944-4_2 detski.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 33 No. 3 2023