Knowledge Graph — Coursera Notes › Organizations › Amazon › Amazon Web Services › Amazon Bedrock › AI agent

Multiagent System

concept · part of AI agent

A multiagent system (MAS) consists of multiple interacting agents, each with individual goals, perceptions, and actions. Agents can collaborate or compete to solve complex tasks, operate in dynamic environments, or simulate real-world systems. Key features include multiple agents, interaction, decentralization, autonomy, and collaboration or competition.

MASs are used in robotics, logistics, simulations, and other domains requiring flexibility, scalability, and robustness.

Each agent in a MAS operates independently, making decisions based on its perception of the environment and internal state without centralized control. Autonomy ensures agents can function even when communication breaks down or other agents fail. Example: autonomous vehicles navigating traffic without direct human control.

Systems can be fully decentralized (e.g., swarm robotics), partially centralized (e.g., sensor networks with a central processor), or fully centralized with agent-like roles (e.g., industrial automation).

Agents interact via direct communication (explicit messages) or indirect communication (shared environments or blackboard systems). Effective communication is key for collaboration or competition. Example: drones sharing locations to maintain safe distances; communication may be peer-to-peer, group-wide, or via a central site for human oversight.

Agents may cooperate to achieve common goals (sharing information, distributing tasks) or compete for limited resources (self-interested, using negotiation or adversarial strategies). Cooperative agents maximize collective benefits; competitive agents optimize individual outcomes. Example: trading bots competing in a financial market simulation.

Coordination manages dependencies between agents' activities to avoid conflicts and redundancy. Key concepts include task allocation (assigning tasks to maximize efficiency) and role assignment (agents take on specific roles based on capabilities). Example: search-and-rescue drones dividing a search area using a grid-search algorithm.

Agents learn from environment and interactions using reinforcement learning, supervised learning, or unsupervised learning. They adapt behavior based on past experiences or environmental changes. Example: game-playing agents adjusting strategies after previous games to counter opponents.

Collective actions of agents produce system-wide outcomes not explicitly programmed but emerging from interactions. This can lead to unpredictable, innovative solutions. Example: flock of birds following simple local rules (maintain distance, align direction) resulting in complex coordinated movement without a leader.

Flexibility allows the system to adapt to new tasks without significant reprogramming. Robustness ensures continued operation despite agent failures or environmental changes. Key concepts: fault tolerance (other agents take over tasks) and scalability (add/remove agents without disruption). Example: delivery drones rerouting if one malfunctions.

Agents are designed to achieve specific objectives, acting rationally by choosing actions that maximize goal achievement. In cooperative systems, they work toward collective goals; in competitive systems, they maximize individual outcomes. Example: self-driving cars aiming to reach destinations efficiently while considering traffic and safety.

In warehouse management, agents (robots) handle tasks like retrieving and restocking items while avoiding conflicts and optimizing workflow.

Inside Multiagent System (1)

• Task allocation in multiagent systems — Tasks can be allocated to agents based on proximity or capability to minimize time and energy, as in a warehouse simulation where the nearest robot is assigned a task using a function that finds the closest robot to the task location.

This is the text view of an interactive 3D knowledge graph — open this page with JavaScript enabled to explore it visually.