AI 용어집
인공지능 완전 사전
Parameter Sharing
Scalability technique where multiple agents share the same neural network parameters, reducing computational complexity and promoting cooperative learning.
Graph Neural Networks for MARL
Network architecture adapted to multi-agent structures where nodes represent agents and edges represent their relationships, enabling efficient communication at large scale.
Multi-Agent Curriculum Learning
Progressive training strategy where task complexity and number of agents gradually increase, improving learning stability at large scale.
Hierarchical MARL
Organizational structure where agents are arranged in hierarchies, allowing management of complex systems with thousands of agents through responsibility decomposition.
Learned Communication Protocols
Mechanisms where agents automatically develop optimized communication protocols to minimize bandwidth while maximizing effective coordination.
Multi-Agent Attention Mechanisms
Technique allowing agents to selectively focus on the most relevant information among thousands of other agents, reducing computational complexity.
Population Based Training for MARL
Evolutionary optimization method where a population of multi-agent policies evolves in parallel, enabling efficient exploration of cooperative strategy space.
QMIX Algorithm
Multi-agent Q-learning algorithm ensuring monotonicity between individual values and joint value, enabling stable learning in large-scale systems.
Opponent Modeling
Ability of agents to model and predict the behavior of other agents, essential for effective coordination in scalable multi-agent systems.
Emergent Communication
Phenomenon where agents spontaneously develop structured communication systems to collectively solve complex problems at large scale.
Agent Modeling Networks
Neural networks specialized in learning the mental models of other agents, crucial for prediction and coordination in massively multi-agent systems.
Distributed MARL Frameworks
Software infrastructures enabling parallelization of multi-agent training on computing clusters, essential for managing millions of agents simultaneously.
Scalable Policy Optimization
Set of algorithmic techniques optimizing the time and space complexity of policy optimization for systems with millions of parameters and agents.
Swarm Intelligence Algorithms
Bio-inspired approaches where simple agents follow local rules to emerge intelligent collective behavior, applicable to very large-scale systems.
Multi-Agent Credit Assignment
Fundamental problem of correctly attributing rewards and penalties to individual agents in a large-scale cooperative system.