KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Sutton Options
Fundamental concept introduced by Sutton defining extended temporal actions with their own policies, entry conditions, and termination functions.
MAXQ Decomposition
Hierarchical decomposition approach based on tasks that structures problems into subtask graphs with value decomposition.
HAM (Hierarchical Abstract Machines)
Hierarchical finite state machines that define abstract policies and sub-policies to guide learning.
FeUdal Networks
Feudal-inspired architecture with managers that define objectives and workers that execute low-level actions.
Goal-Conditioned RL
Reinforcement learning where policies are conditioned by sub-goals to facilitate hierarchical decomposition.
Subgoal Discovery
Automatic techniques for identifying and discovering relevant subgoals in the state space without human supervision.
Temporal Abstraction
Methods for abstracting decisions across different time scales to manage long temporal horizons.
Multi-task HRL
Simultaneous hierarchical learning on multiple tasks sharing common subtasks for efficient transfer.
Intrinsic Motivation in HRL
Use of intrinsic rewards to guide the automatic discovery of relevant hierarchical structures.
Meta-learning in HRL
Meta-learning approaches to automatically adapt hierarchical structures to new tasks and environments.
Option Discovery Methods
Specific algorithms for automatically discovering effective options based on density, reachability, or bottleneck.
Hierarchical Policy Gradient
Methods of policy gradient adapted to hierarchical structures with simultaneous optimization of hierarchical levels.
State Abstraction in HRL
State abstraction techniques to simplify representations at different hierarchical levels and accelerate learning.
Termination Functions
Functions deciding when to terminate options and subtasks, crucial for the efficiency of hierarchies.
Hierarchical Actor-Critic
Hierarchical actor-critic architecture with coordinated multi-level actors and critics for hierarchical learning.