Glosarium AI
Kamus lengkap Kecerdasan Buatan
State-action distribution
Probabilistic representation of the Q(s,a) value function that models the complete distribution of possible returns rather than just their mathematical expectation.
Distributional transition model
Model-based reinforcement learning model that captures uncertainty in state transitions by modeling probability distributions over next states.
Probabilistic dynamics model
Predictive model in model-based RL that generates probability distributions over next states or rewards rather than deterministic predictions.
Epistemic uncertainty in RL
Uncertainty due to lack of knowledge about the environment model, modeled by distributions in distributional model-based RL approaches.
Aleatoric uncertainty in RL
Inherent uncertainty in the environment that cannot be reduced even with more data, captured by distributions in distributional RL models.
Distributional policy gradient
Extension of policy gradient methods that directly optimizes over the distribution of returns rather than their expectation, enabling risk-sensitive policies.
Risk-sensitive RL
Reinforcement learning approach that uses distributional information to optimize risk metrics like CVaR or standard deviation instead of just expectation.
Model ensembles in distributional RL
Technique using multiple independently learned models to capture epistemic uncertainty in distributional model-based RL approaches.
Particle-based distribution models
Distributional modeling approach that represents distributions by a set of weighted particles, useful for complex transitions in model-based RL.
Wasserstein distance in distributional RL
Metric used to measure dissimilarity between distributions in the distributional Bellman operator, offering better convergence properties than KL distance.
Moment matching in distributional RL
Optimization technique that adjusts parameters to match statistical moments (mean, variance, etc.) of predicted and target distributions.
Variational inference in RL
Method for approximating complex distributions by optimizing a family of simpler distributions, applied in model-based RL to handle uncertainty.
Bayesian model-based RL
Approach that maintains a distribution over possible environment models, using Bayesian methods to quantify and exploit epistemic uncertainty.
Distributional Bellman operator
Extension of the classic Bellman operator that operates on return distributions rather than scalar values, preserving distributional structure.
Horizon-dependent distributions
Concept in distributional RL where the return distribution changes with the time horizon, capturing the evolution of uncertainty over different time scales.
Categorical atomic projection
Mathematical operation used in C51 that projects the target distribution onto predefined atom support to maintain distributional consistency.
Distributional uncertainty propagation
Process in model-based RL where the uncertainty of model predictions is propagated through planning steps to evaluate policy robustness.