AI-woordenlijst
Het complete woordenboek van kunstmatige intelligentie
Bootstrap in RL
Resampling technique used in reinforcement learning to estimate value function uncertainty by creating multiple estimations from the same data sample.
Bootstrap Value Distribution
Probabilistic representation of the value function obtained by aggregating multiple bootstrap estimations, allowing quantification of uncertainty on value predictions.
Weighted Bootstrap
Technique assigning weights to bootstrap samples based on their relevance or recency to give more importance to more informative experiences in value estimation.
Q-learning with Bootstrap
Extension of classic Q-learning using multiple Q-value heads trained on different bootstrap samples to capture uncertainty and improve exploration.
C51 (Categorical 51)
Distributional algorithm discretizing the return distribution into 51 probability atoms, using bootstrap techniques to estimate uncertainty on this representation.
IQN (Implicit Quantile Networks)
Network architecture directly learning the quantile distribution of returns, integrating bootstrap mechanisms to quantify uncertainty of quantile predictions.
QR-DQN (Quantile Regression DQN)
DQN variant using quantile regression on bootstrap samples to learn the complete distribution of action values with uncertainty quantification.
Bootstrap Head Networks
Architecture comprising multiple independent output heads trained on different bootstrap samples to capture uncertainty in value predictions.
Uncertainty-based Exploration
Exploration strategy using bootstrap estimates to quantify uncertainty and guide the agent toward the least known states of the environment.
Bootstrap Ensembles
Method training multiple models on different bootstrap samples to form a predictive ensemble capturing the variability and uncertainty of the learning process.
Dropout as Bootstrap Approximation
Technique using dropout during inference as an efficient approximation of bootstrap to quickly estimate uncertainty without training multiple models.
Credible Intervals
Statistical intervals derived from bootstrap distributions quantifying uncertainty on value estimates with a specified confidence probability.
Bootstrap Variance
Metric quantifying the dispersion of bootstrap estimates among themselves, serving as a direct indicator of epistemic uncertainty in value predictions.
Bootstrap Bias
Systematic deviation potentially introduced by bootstrap methods, requiring correction techniques such as double bootstrap for unbiased estimates.
Sequential Bootstrap
Variant adapted to temporal RL data preserving sequential dependency structure during resampling to avoid underestimation of uncertainty.