YZ Sözlüğü
Yapay Zekanın tam sözlüğü
Hierarchical Reinforcement Learning
Learning paradigm where decision policies are structured in hierarchical levels, allowing complex tasks to be decomposed into simpler and reusable sub-tasks.
Sutton's Options
Extended temporal action units that combine sequences of atomic actions into reusable macroscopic behaviors, forming the basis of temporal abstraction in hierarchical RL.
Task Decomposition
Algorithmic process of automatic segmentation of complex objectives into hierarchically organized sub-objectives to facilitate learning and optimization.
Hierarchical Policies
Set of decision policies organized in layers where high-level policies select sub-tasks and low-level policies execute the corresponding actions.
Temporal Abstraction
Technique grouping primitive actions into coherent temporal sequences, reducing planning complexity and improving learning efficiency.
Hierarchical Meta-Learning
Approach where the system learns to learn optimal hierarchical structures, adapting quickly to new tasks by reusing acquired meta-knowledge.
Weight Consolidation
Mechanism protecting important synaptic weights for previous tasks, typically via regularization penalties, to prevent forgetting during new learning.
Hierarchical Replay Buffer
Hierarchically organized data structure storing and selectively reusing past experiences to maintain skills while learning new tasks.
Task Graph
Formal representation of dependencies and relationships between sub-tasks, guiding the automatic construction of optimal policy hierarchies.
Hierarchical Transfer Learning
Selective transfer of knowledge between hierarchical levels, enabling the reuse of effective sub-policies to accelerate learning of new complex tasks.
Continual Learning Stabilization
Set of algorithmic techniques ensuring stable convergence of models during sequential acquisition of skills, preventing oscillations and divergence.
Reusable Sub-Policies
Atomic decision modules trained independently that can be dynamically combined to form complex policies, promoting modularity and efficiency.
Multi-Timescale Learning
Framework integrating simultaneous decisions at different temporal horizons, from immediate actions to long-term strategies, for optimal complexity management.