KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Active Reinforcement Learning
Hybrid methodology combining active learning and reinforcement learning principles to optimize sample selection for annotation.
Sample Selection Policy
Deterministic or stochastic strategy defining which data to request for annotation to maximize model improvement under budget constraints.
Reinforcement Learning Agent
Algorithmic entity that learns to make optimal sample selection decisions through interaction with the annotation environment.
Reward Function
Signal quantifying the utility of each sample selection action, typically based on model performance improvement.
State-Action-Value
Q(s,a) function estimating the expected cumulative reward when selecting action a from state s and following the optimal policy.
Deep Reinforcement Learning
Extension of reinforcement learning using deep neural networks to approximate value functions or policies.
Uncertainty-Based Active Learning
Strategy where the agent preferentially selects samples for which the model exhibits the highest predictive uncertainty.
Strategic Sample Selection
Optimized decision-making process aiming to identify data subsets maximizing information gain per annotation cost.
Off-Policy Reinforcement Learning
Method enabling the learning of an optimal policy while following a different behavior policy, useful for flexible exploration.
Online Reinforcement Learning
Paradigm where the agent learns and selects samples simultaneously during annotation, dynamically adapting its strategy.
Learning-Annotation Trade-off
Optimization of the balance between time spent on intelligent selection and potential gains in model performance.
Data Acquisition Strategy
Systematic action plan for identifying and collecting the most relevant data to annotate according to predefined criteria.
Multi-Agent Reinforcement Learning
Extension where multiple agents collaborate or compete to jointly optimize the sample selection strategy.
Active Q-Learning Algorithm
Variant of Q-learning adapted to active learning, where actions correspond to selecting samples to annotate.
Guided Exploration Policy
Exploration strategy oriented towards regions of the data space potentially most informative for the model.
Bayesian Reinforcement Learning
Method integrating uncertainty into value function estimation for more robust decision-making in sample selection.