Słownik AI
Kompletny słownik sztucznej inteligencji
Multi-Step Distributional TD
Temporal-difference algorithm that propagates information over multiple time steps in the distribution space, improving the stability and efficiency of learning.
Quantile Regression in RL
Distributional approach that directly estimates the quantiles of the return distribution, offering a flexible representation without requiring prior discretization.
Wasserstein Metric
Distance between distributions used in distributional learning to measure the similarity between return distributions, taking into account the geometry of the reward space.
N-Step Return Distribution
Probability distribution of the sum of rewards over N future steps, used to accelerate information propagation in multi-step distributional algorithms.
Distributional Policy Evaluation
Process of estimating the complete return distribution for a given policy, rather than just its expected value, allowing for finer performance analysis.
Risk-Sensitive RL
Extension of distributional reinforcement learning that optimizes specific risk measures (CVaR, variance) rather than expectation alone.
Distributional Policy Gradient
Policy optimization algorithm that uses the complete information of the return distribution to update parameters, enabling explicit risk-reward trade-offs.
Distributional Actor-Critic
Architecture where the critic evaluates the return distribution rather than a single scalar value, providing a richer learning signal to the actor.
Distributional Dynamic Programming
Extension of dynamic programming methods that operates on value distributions, allowing more precise resolution of problems with uncertainty.
Atomic Support in C51
Discrete set of predefined values used as support to represent return distributions in the C51 algorithm, allowing efficient approximation of continuous distributions.
Distributional Bootstrap
Estimation technique where the distribution of a state is updated using the distribution of next states, preserving the stochastic structure across iterations.
Stability in Distributional RL
Property guaranteeing the convergence of distributional algorithms, often improved through the use of multi-step methods and appropriate projections.
Distributional Risk Measures
Functionals of the return distribution (Value-at-Risk, Expected Shortfall) used to characterize and optimize behavior in the face of uncertainty.
Multi-Step Uncertainty Propagation
Mechanism by which uncertainty about future returns is effectively propagated across multiple time horizons in the distributional framework.
Distributional Sampling
Sampling technique from predicted return distributions to estimate gradients and update policies in distributional algorithms.