Multi-agent technology is a new way of solving complex problems, using the principles of self-organization and evolution, inherent in living systems.

The essence of multi-agent technology lies in a fundamentally new method for solving complex problems, which cannot be solved or are difficult to solve by classical mathematical methods.

In contrast to the classical method of solving problems, when a combinatorial search for solutions is performed on a clearly defined (deterministic) algorithm in order to find the best solution to the problem, in multi-agent technology the solution is obtained in the course of self-organization of a multitude of software agents capable of competition and cooperation and having their own criteria, preferences and limitations. The solution is found when, after nondeterministic interactions, agents achieve an unimprovable consensus (temporary equilibrium or balance of interests), which is taken as the final solution of the problem.

The solution in such systems is always regarded as a temporary «equilibrium» (unstable equilibrium or stable nonequilibrium), obtained as a dynamic shutdown of the system when none of the agents can improve their state. It indicates achievement of a reasonable compromise, balance of interests or harmony of all participants, even if some of the agents are not fully satisfied (they simply do not have other better options).

Agents can act on behalf of both a person, and any physical or abstract entities, as it is planned within the Internet of Things, in order to take into account the action and find the balance of as many factors as possible.

There are many definitions of an agent, but the main features of a software agent are as follows:

  • autonomy, i.e. it is capable of setting and achieving the goal;
  • reaction to changes in the environment, making decisions and executing them to achieve the goal;
  • sociality: proactively interacts with other agents or users

In our understanding, the main distinguishing features of multi-agent technology can be shown in the diagram below:

Traditional systems vs MAS

In a multi-agent model, for each entity of the real world there is a software agent representing the interests of the given entity and coordinating its decisions with other agents.

The advantages of multi-agent technology that allow for building self-organizing systems, are especially conspicuous in the conditions of a priori uncertainty and high dynamics of the surrounding world, making it possible to create adaptive systems that rebuild their plans in accordance with events in real time.

Thus, in the classical methods of planning and optimization, it is considered that all orders and resources are known in advance and do not change in the course of problem solving.  Besides, dimension of the problem is considerably limited in order to avoid a combinatorial explosion and exponentially fast slowdown of problem solution.

In the proposed models, methods and algorithms, a Distributed Problem Solving approach is initially used: a complex task is broken down into many small ones, and then conflicts between solutions are resolved through self-organization. At the same time, the system does not seek for a single global solution, but through multiple parallel and asynchronous interactions, quickly finds an acceptable rational solution, despite the existence of a multitude of very different and often contradictory criteria. This works for problems of any dimension.

A step towards AI: swarm intelligence and emergent intelligence

We are used to the fact that computers always operate strictly according to the given program, which has nothing in common with human intelligence.

In our opinion, human intelligence is built according to entirely different principles, as a self-organizing non-equilibrium thermodynamic system, which makes it possible to navigate in a complex situation, deal with indeterminate problems, adapt to changing conditions, etc.

In this context, multi-agent technology offers new models, methods and tools for creating truly intelligent systems, capable of solving complex problems independently in conditions of uncertainty and high dynamics of changes.

To solve complex resource management problems, we propose a set of special demand and resource agents that interact in the virtual market of the system and build relations among themselves, forming a demand and resource network (DR-network). These agents maximize their own satisfaction functions considering the given functions of bonuses and fines.

Demand-Supply Match

The system constructed in this way demonstrates the role of an intelligent resonator, which makes it possible to produce rather complex solutions even in case of relatively simple agents and small changes at the input. Such solutions emerge as a result of long chains of autocatalytic reactions with revisions of earlier decisions.

In this case, we can mention the phenomenon of Swarm Intelligence — as an important alternative to the classical understanding of an intelligent system, which is now accepted in artificial intelligence (AI). That is, a system that is mechanically assembled with such components as induction and deduction units, etc.

Indeed, mental capabilities of one ant or a bee are relatively small, however, acting together as a single organism, a swarm of bees or a colony of ants is a powerful force with a high degree of intelligence, enabling them to protect the nest from unforeseen invasions, constantly explore new territories, find food in an unfamiliar area and solve many other crucial life tasks in constantly changing conditions of the surrounding world.

To develop «swarm intelligence», one can create models of increasingly complex team interactions, including new classes of agents and protocols of their negotiations to achieve concessions, learn from experience, etc.

The higher the intelligence of each agent and the richer the possibility of such communication between them — the more complex and creative the demonstrated behavior of the system.

Such systems, by definition, inherently have a completely different phenomenology associated with nondeterministic behavior, phenomena of order and chaos, bifurcations, catastrophes and many other nonlinearities.

Different classes of such models of the new-generation AI will be called «Emergent Intelligence» (EI), reflecting the inherent nature of self-organization.

In comparison with AI, there is no main control unit in EI, which could be responsible for the system intelligence. On the contrary, EI is viewed as a temporarily emerging property of a self-organizing system.

For an observer, EI can display itself as an autocatalytic reaction or a chain of coordinated changes in the system of agents’ decisions. This reaction arises spontaneously, at a previously unknown time, and propagates in the system as a wave of coordination and approval (like a fire in a forest or a lightning during a storm). After that, it unexpectedly disappears, but during its existence, it determines the work of its elements.

As a result, it may seem that EI arises «out of thin air», but in fact, it appears due to the potential energy of earlier decisions reflecting accumulated dissatisfaction or nonequilibrium. However, in the course of its existence, it decisively «rules» the operation of the entire system, just like a traffic jam on the road controls all the drivers.

This phenomenon of «double helix» in decision-making is known in the theory of self-organization, where local interactions of agents form global structures, which, in their turn, influence behavior of local agents that have formed them (the Kaufmann principle).

A great contribution to development of this research area was made by Alexander Bogdanov (organization theory), Ilya Prigogine (self-organization in physical systems), Marvin Minsky (psychology and theory of mind), Arthur Koestler (biology) and a number of other scientists.

Currently, multi-agent technology is one of the most dynamically developing and promising directions in the field of information technologies, successfully complementing such advanced areas as semantic Internet and ontologies, network-centric systems, Internet of Things and others.

According to Gartner, the world-famous company, in their evaluation of information technology market, multi-agent technology will be the basis for more than 40% of all mobile applications by 2020.

Up-and-coming application fields for multi-agent technology:

Currently, the following promising areas of application of multi-agent technology are distinguished:

  • Industry
  • Transport
  • Power industry
  • Supply Chain
  • E-commerce
  • Intelligent search for goods and services on the Internet
  • Targeted advertising and marketing
  • Warfare
  • Healthcare
  • Building
  • Communication

In these areas, the following problems can be solved:

  • Resource management
  • Construction of complex products
  • Designing
  • Monitoring and control
  • Image recognition
  • Understanding of texts
  • Knowledge acquisition

Important prospects of this technology are associated with development of the Internet of Things and ubiquitous computing.

Results of application of multi-agent technology:

Development of intelligent systems based on multi-agent technology makes it possible to achieve the following results:

  • Solving complex tasks that could not previously be automated;
  • Results of such solutions provide such high quality that is comparable to decisions made by people;
  • The initial solution is built efficiently (linearly or polynomially);
  • Changes in problem statement lead only to adaptation of the solution «on the fly»;
  • Taking into account events in real time is supported;
  • It is possible to solve the problem in a dialogue with the user;
  • Calculations can be easily paralleled to solve extra-complex tasks.

As a result, multi-agent technology makes it possible to build intelligent systems of a new generation, characterized by high openness, flexibility and efficiency, productivity, scalability, reliability and survivability.