AI 용어집
인공지능 완전 사전
Warm-up chain
Initial sequence of MCMC iterations used to allow the Markov chain to reach its stationary distribution before collecting samples for inference.
Burn-in phase
Initial period of the MCMC algorithm during which samples are discarded because they do not yet come from the target stationary distribution.
Convergence period
Time interval required for the Markov chain to asymptotically reach its stationary distribution with a given precision.
Adaptive sampling
Technique where MCMC algorithm parameters are dynamically adjusted during the warm-up phase to optimize sampling efficiency.
Parameter tuning
Process of adjusting MCMC algorithm hyperparameters (step size, variances) during the warm-up phase to improve the acceptance rate.
Burn-in length
Number of initial iterations to discard in an MCMC chain, determined empirically or via convergence diagnostics.
Convergence diagnostic
Set of statistical methods (Gelman-Rubin, Raftery-Lewis) used to evaluate whether the chain has reached its stationary distribution.
Scaling samples
First generated samples used to estimate the characteristics of the target distribution and adjust the chain parameters.
Warm-up chain
Initial Markov chain whose sole purpose is to reach thermodynamic equilibrium before starting the main sampling.
Chain stabilization
Process by which the Markov chain loses memory of its initial state and converges to its invariant distribution.
Starting points
Initial values chosen to initialize the MCMC chain, whose influence must disappear after the warm-up phase.
Transition samples
Intermediate observations generated during the warm-up phase that are not used in the final inference.
Warm-up time
Computational duration required for the chain to reach thermodynamic equilibrium, measured in number of iterations or CPU time.
Detailed balance
Mathematical condition ensuring that the probability flow between two states is reversible, essential for MCMC convergence.
Acceptance-rejection
Fundamental mechanism of MCMC algorithms where proposals are accepted or rejected according to a probabilistic criterion.
Effective sample size
Number of equivalent independent samples in an autocorrelated chain, calculated after discarding burn-in.