AI 용어집
인공지능 완전 사전
Continuous Quantile Distribution
Representation of the return distribution as a set of evolving quantiles in continuous action spaces, allowing fine modeling of uncertainty and risks.
Cramer-Wold Distributional Projection
Mathematical technique enabling comparison of distributions by projecting onto one-dimensional directions, essential for distributional metrics in continuous RL.
Atomic Distribution Network
Neural architecture representing a distribution as a weighted set of fixed atoms, suitable for continuous action problems with stochastic returns.
Distributional Risk in Continuum
Measure quantifying uncertainty in return distributions of continuous action spaces, crucial for robust policy evaluation.
Distributional Stochastic Policy
Action strategy directly incorporating return distribution in continuous action selection, optimizing over the entire distribution rather than just the expectation.
Quantile Distribution Expectation
Operator calculating expectation from quantile representation, preserving distributional properties in continuous spaces.
Distributional Rejection Sampling
Sampling method preserving distributional properties when generating continuous actions from complex return distributions.
Stochastic Distributional Optimization
Optimization paradigm working directly on return distributions rather than their point estimates in continuous spaces.
Distributional Kernel Approximation
Technique using kernel functions to approximate return distributions in high-dimensional continuous action spaces.
Wasserstein Distance in Continuous RL
Metric measuring the dissimilarity between return distributions, particularly adapted to continuous action problems with complex geometry.
Distributional Importance Sampling
Weighted sampling technique preserving distributional characteristics when estimating policy gradients in continuous settings.
Distributional Monte-Carlo Update
Algorithm updating return distributions using Monte-Carlo samples in continuous action spaces, preserving the distributional shape.
Distributional Variance Reduction
Set of techniques aiming to reduce variance in distributional estimates without losing information about distribution shapes.
Distributional Greedy Policy
Strategy selecting optimal actions based on criteria on the full distribution (e.g., quantile, CVaR) rather than just expectation in continuous settings.
Distributional Bellman Equation
Formulation of the Bellman equation operating on complete distributions rather than scalar values, fundamental to continuous distributional RL.
Continuous Distributional Critic
Neural network estimating the complete return distribution for continuous state-action pairs, replacing the traditional scalar value critic.
Distributional Bias in Continuous Action
Phenomenon where distributional approximations introduce systematic biases in return estimation in continuous action spaces.
Continuous Distributional Normalization
Normalization technique preserving distributional properties when processing continuous actions at different scales.
Adaptive Distributional Exploration
Exploration strategy using complete return distribution information to adapt exploratory behavior in continuous action.