YZ Sözlüğü
Yapay Zekanın tam sözlüğü
Return distribution
Complete probabilistic representation of the sum of discounted future rewards, capturing all possible scenarios rather than a single expected value.
Quantile distribution
Approach that directly models the quantiles of the return distribution to capture the variability and distribution tails of rewards.
Conditional value at risk
Robust risk measure calculating the expected value of returns in the lower tail of the distribution, beyond a specified quantile.
Implicit distribution
Distributional representation learned indirectly without explicit parameters, often through generative neural networks or samplers.
Return variance
Dispersion measure quantifying the mean square deviation of returns from their expectation, a key indicator of risk in decisions.
Policy entropy
Uncertainty measure on the agent's actions, used to explore the state-action space and quantify behavioral uncertainty.
Confidence bound
Statistical intervals guaranteeing with a predefined probability that the true value lies within the estimated range, essential for safe exploration.
Cramer distribution
Family of flexible distributions allowing modeling of skewness and heavy tails in returns, beyond Gaussian assumptions.
Kernel estimation
Non-parametric method for estimating the probability density of returns using kernel functions to smooth empirical observations.
Uncertainty propagation
Process of propagating uncertainty through successive steps of reinforcement learning, from observations to final decisions.
Variational approximation
Optimization method approximating complex distributions by simpler families, minimizing divergence between distributions.
Mixture distribution
Weighted combination of several base distributions, allowing to capture multimodal behaviors in expected returns.
Cumulative distribution function
Function F(x) giving the probability that the return is less than or equal to x, completely characterizing the distribution of returns.
Bias-variance tradeoff
Fundamental dilemma between model complexity (high variance, low bias) and its simplicity (low variance, high bias) in distributional estimation.