YZ Sözlüğü
Yapay Zekanın tam sözlüğü
POMDP (Partially Observable Markov Decision Process)
Theoretical framework modeling environments where the agent perceives only a partial observation of the true state, requiring probabilistic inference about the hidden state to make optimal decisions.
Observation Space
Set of partial sensory signals that each agent can perceive from the environment, representing incomplete information about the global state of the system.
Belief State
Probability distribution over the hidden state space that an agent maintains and updates from its successive observations to represent its uncertainty about the true state of the environment.
Communication Protocol
Mechanism defining when, how, and what information agents can exchange among themselves to coordinate their actions in a partially observable environment.
Centralized Training with Decentralized Execution
Approach where agents train using global information (states, actions of all agents) but execute their policies individually using only their local observations.
Value Function Factorization
Technique decomposing the global value function into a sum of individual or local value functions, enabling decentralized learning while preserving global consistency.
Adversary Modeling
Process of inferring the policies or intentions of other agents based on their observed behaviors, crucial for decision-making in competitive or cooperative environments.
Credit Assignment Problem
Difficulty in correctly attributing the global reward to each agent in a multi-agent system, particularly complex when observations are partial and actions are interdependent.
Joint Action Learning
Method where agents learn to coordinate their actions by explicitly modeling the impact of combined actions on the global reward, despite partial observability.
State Estimation
Algorithmic process allowing an agent to infer the most probable global state from its local observations and its model of the environment.
Information Sharing
Strategy defining how agents distribute and aggregate their local observations to improve the collective knowledge of the environment's state.
Local Observation History
Temporal sequence of an agent's past observations, used as additional context to compensate for the lack of information about the current global state.
Multi-agent Partial Observability
Condition where no individual agent can observe the complete state of the system, requiring coordination and inference strategies to achieve optimal performance.
Decentralized Policy
Decision function for each agent that maps its local observation history to an action, without direct dependence on other agents' information during execution.
Common Knowledge
Information that all agents know and know that others also know, essential for coordination in partially observable environments.
Coordination Graph
Structure representing interaction dependencies between agents, allowing the global decision problem to be factored into easier-to-solve local subproblems.