A multi-agent system (MAS) or "self-organized system" is a computational system composed of multiple interacting intelligent agents. Multi-agent systems can solve problems that are difficult or impossible for an individual agent or a monolithic system to solve. Intelligence may include methodic, functional, procedural approaches, algorithmic search or reinforcement learning. With advancements in large language models (LLMs), LLM-based multi-agent systems have emerged as a new area of research, enabling more sophisticated interactions and coordination among agents. Despite considerable overlap, a multi-agent system is not always the same as an agent-based model (ABM). The goal of an ABM is to search for explanatory insight into the collective behavior of agents (which do not necessarily need to be "intelligent") obeying simple rules, typically in natural systems, rather than in solving specific practical or engineering problems. The terminology of ABM tends to be used more often in the science, and MAS in engineering and technology. Applications where multi-agent systems research may deliver an appropriate approach include online trading, disaster response, target surveillance and social structure modelling.
Concept Multi-agent systems consist of agents and their environment. Typically, research on multi-agent systems refers to software agents. However, the agents in a multi-agent system could equally well be robots, humans, or human teams, and may consist of combined human-agent teams. Agents can be divided into types spanning simple to complex. Categories include:
Passive agents or "agent without goals" (such as obstacle, apple or key in any simple simulation) Active agents with simple goals (like birds in flocking, or wolf–sheep in prey-predator model) Cognitive agents, with beliefs, desires, intentions, and commitments processed by logical, probabilistic, and neural network-based reasoning Agent environments can be divided into:
Virtual Discrete Continuous Agent environments can also be organized according to properties such as accessibility (whether it is possible to gather complete information about the environment), determinism (whether an action causes a definite effect), dynamics (how many entities influence the environment in the moment), discreteness (whether the number of possible actions in the environment is finite), episodicity (whether agent actions in certain time periods influence other periods), and dimensionality (whether spatial characteristics are important factors of the environment and the agent considers space in its decision making). Agent actions are typically mediated via an appropriate middleware. This middleware offers a first-class design abstraction for multi-agent systems, providing means to govern resource access and agent coordination.
Characteristics The agents in a multi-agent system have several important characteristics:
Autonomy: agents are at least partially independent, self-aware, autonomous Local views: no agent has a full global view, or the system is too complex for an agent to exploit such knowledge Decentralization: no agent is designated as controlling (or the system is effectively reduced to a monolithic system)
Self-organisation and self-direction Multi-agent systems can manifest self-organisation as well as self-direction and other control paradigms and related complex behaviors even when the individual strategies of all their agents are simple. When agents can share knowledge using any agreed language, within the constraints of the system's communication protocol, the approach may lead to a common improvement. Example languages are Knowledge Query Manipulation Language (KQML) or Agent Communication Language (ACL).
Decision-Making Decision protocols in multi-agent systems refer to the structured rules and procedures that agents follow to reach collective decisions or agreements. Such protocols specify how agents share information, negotiate, and resolve conflicts, ensuring coordinated behavior and effective joint actions. Decision protocols can range from voting mechanisms to consensus-building algorithms, and they significantly influence the efficiency and reliability of multi-agent interactions.
System paradigms Many MAS are implemented in computer simulations, stepping the system through discrete "time steps". The MAS components communicate typically using a weighted request matrix, e.g.
Speed-VERY_IMPORTANT: min=45 mph, Path length-MEDIUM_IMPORTANCE: max=60 expectedMax=40, Max-Weight-UNIMPORTANT Contract Priority-REGULAR
and a weighted response matrix, e.g.
Speed-min:50 but only if weather sunny, Path length:25 for sunny / 46 for rainy Contract Priority-REGULAR note – ambulance will override this priority and you'll have to wait
A challenge-response-contract scheme is common in MAS systems, where
First a "Who can?" question is distributed. Only the relevant components respond: "I can, at this price". Finally, a contract is set up, usually in several short communication steps between sides, also considering other components, evolving "contracts" and the restriction sets of the component algorithms. Another paradigm commonly used with MAS is the "pheromone", where components leave information for other nearby components. These pheromones may evaporate/concentrate with time, that is their values may decrease (or increase).
Properties MAS tend to find the best solution for their problems without intervention. There is high similarity here to physical phenomena, such as energy minimizing, where physical objects tend to reach the lowest energy possible within the physically constrained world. For example: many of the cars entering a metropolis in the morning will be available for leaving that same metropolis in the evening. The systems also tend to prevent propagation of faults, self-recover and be fault tolerant, mainly due to the redundancy of components.
Research The study of multi-agent systems is "concerned with the development and analysis of sophisticated AI problem-solving and control architectures for both single-agent and multiple-agent systems." Research topics include:
… excerpt ends here. Continue reading the full article.
