Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
FeUdal Network (FuN)
Hierarchical reinforcement learning architecture inspired by the feudal system where a manager defines long-term goals and workers execute concrete actions. This structure enables better generalization and efficient management of complex state spaces.
Manager Module
High-level component in the FuN architecture that generates sub-goals or abstract directions to guide workers. Operates at a slower time scale and focuses on strategic planning rather than execution details.
Worker Module
Lower-level agent in the FuN hierarchy responsible for executing primitive actions by following manager directives. Transforms abstract goals into concrete actions in the environment at a higher frequency.
Goal Representation
Vector encoding of sub-goals generated by the manager and communicated to workers in the FuN architecture. Enables efficient communication between hierarchical levels and facilitates learning of goal-conditioned policies.
Space Factorization
Technique in FuN involving the decomposition of state space into distinct management and exploitation spaces. Allows the manager to focus on global aspects while workers handle local environment details.
Goal-Conditioned Policy
Worker action policy in FuN that considers both the current state and the goal provided by the manager. Enables flexible action execution based on received hierarchical directives.
End-to-End Training
Simultaneous learning process of managers and workers in FuN through gradient backpropagation across the entire hierarchy. Ensures consistent optimization of the complete architecture without requiring separate training for each component.
LSTM Manager
Manager implementation using LSTM networks to maintain memory of past goals and plan long-term. Enables consistent management of sequential sub-goals in complex tasks.