Słownik AI
Kompletny słownik sztucznej inteligencji
Multi-level Policy Optimization
Coordinated optimization process adjusting parameters of multiple hierarchical policy layers using synchronized gradients to maximize overall reward.
Nested Policy Networks
Neural network architecture where low-level policies are nested within high-level policies, enabling hierarchical decomposition of decisions and actions.
Option Framework
Mathematical formalization of extended temporal behaviors in hierarchies, where each option combines an intra-option policy, initiation condition, and termination condition.
Sub-policy Selection
Mechanism by which the high-level policy dynamically selects which sub-policy to activate based on the current state and objectives to be achieved.
Primitive Actions
Fundamental lowest-level actions executed directly in the environment, constituting the building blocks of complex behaviors constructed by the hierarchy.
Hierarchical Advantage Estimation
Advantage estimation technique accounting for the hierarchical structure, evaluating each level's contribution to overall performance improvement.
Cross-level Gradient Flow
Gradient propagation mechanism across different hierarchical levels, ensuring coordinated and stable optimization of the entire architecture.
Hierarchical Entropy Regularization
Regularization technique applying differentiated entropy penalties according to hierarchical levels to balance exploration and exploitation at each scale.
Multi-timescale Learning
Learning paradigm where different hierarchical levels operate at distinct temporal scales, enabling efficient management of short and long-term decisions.
Hierarchical Value Functions
Hierarchically structured value functions estimating expected returns at different levels of temporal abstraction to guide policy learning.