AI 용어집
인공지능 완전 사전
Ensemble of Models
An approach involving training multiple predictive models simultaneously to combine their predictions and reduce generalization error. This technique improves robustness and estimates epistemic uncertainty in reinforcement learning systems.
Ensemble Bootstrapping
A sampling method with replacement used to create varied training datasets, enabling the training of multiple models on different subsets to capture prediction variability. Particularly effective for uncertainty estimation in Model-Based RL.
Ensemble Prediction
A technique that aggregates the predictions of several models forming an ensemble, typically through averaging or voting, to obtain a more stable and accurate final prediction. The variance between individual predictions quantifies the model's uncertainty.
Decision Robustness
The ability of an agent to maintain acceptable performance in the face of model uncertainties and environmental variations, using ensembles to assess decision confidence. Critical for deploying RL agents in real-world environments.
Ensemble Averaging
An aggregation method where the final prediction is the average of the individual predictions from each model in the ensemble, reducing bias and variance while providing a natural measure of uncertainty. The foundation of modern ensemble approaches in RL.
Uncertainty Weighting
A strategy using the uncertainty estimated by ensembles to weight decisions, favoring actions with more certain predictions during exploitation and exploring areas of high uncertainty. Improves the exploration-exploitation trade-off in RL.
Uncertainty-Driven Exploration
An exploration policy that uses ensemble uncertainty measures to guide the agent towards less-known states, optimizing information collection to improve model learning. An effective alternative to curiosity-based exploration methods.
Approximate Bayesian
An approximation of exact Bayesian inference using ensembles of neural networks to estimate the posterior distribution of the model's parameters. Provides a practical probabilistic interpretation for uncertainty quantification in RL.
Aleatoric Uncertainty
Inherent uncertainty in the process, irreducible even with infinite data, resulting from stochastic noise in the environment or observations. Differentiated from epistemic uncertainty in modern quantification approaches.
Ensemble Variance
Metric quantifying the dispersion of predictions between different models in an ensemble, serving as a direct proxy for epistemic uncertainty in Model-Based RL systems. Higher in less explored regions of the state space.
Posterior Predictive Distribution
Complete distribution over future states or rewards incorporating both uncertainty about model parameters and process noise, approximated by ensemble predictions in practice. Fundamental for robust planning in RL.
Sample Efficiency
Measure of an algorithm's ability to learn with minimal interactions with the environment, improved by ensembles that enable efficient knowledge transfer and targeted exploration. Critical for data-intensive RL applications.
Ensemble Generalization
Ability of ensemble methods to better generalize to unseen states by combining knowledge from multiple partially correct models, reducing overfitting and improving robustness to distribution variations.
Ensemble Hyperparameters
Parameters controlling the ensemble configuration, including the number of models, bootstrap rates, aggregation methods, and diversification strategies. Crucial for optimizing the trade-off between performance and computational complexity.