AI 용어집
인공지능 완전 사전
Curiosity-Driven RL
Reinforcement learning approach where the agent generates intrinsic rewards based on its curiosity to encourage the exploration of complex environments with sparse extrinsic rewards.
Intrinsic Motivation
Computational psychological mechanism that drives an agent to act to satisfy internal needs such as curiosity, rather than for task-specific external rewards.
Prediction Error
Measure of the difference between a world model's predictions and the actual observations, used as a curiosity signal to encourage the exploration of unexpected states.
Intrinsic Curiosity Module (ICM)
Neural architecture composed of forward and inverse dynamics models that generates intrinsic rewards based on prediction uncertainty to guide exploration.
Random Network Distillation (RND)
Exploration method where a fixed random neural network is used as a target for a predictor network, with the prediction error serving as an intrinsic reward for novel states.
Count-Based Exploration
Exploration strategy that assigns curiosity bonuses inversely proportional to the visitation frequency of states, thus encouraging the discovery of less explored regions.
Pseudo-counts
Approximate estimation of state visitation frequencies in continuous or high-dimensional spaces, used to implement count-based curiosity bonuses.
Empowerment
Information-theoretic measure quantifying the control an agent exerts over its environment, which is maximized to encourage exploratory behaviors that increase the agent's influence.
Information Gain
Amount of new information acquired by the agent about the environment, used as an intrinsic signal to direct exploration toward the most informative regions.
Episodic Curiosity
Curiosity approach based on short-term memory where the agent is motivated to visit states different from those recently observed in the current episode.
Variational Information Maximization Exploration (VIME)
Exploration method that maximizes mutual information between model parameters and future observations, using Bayesian approaches to quantify uncertainty.
State Visitation Count
Counter of the number of times a particular state has been visited, used to calculate exploration bonuses that favor the discovery of rare or unexplored states.
Curiosity-Driven Exploration
Exploration paradigm where the agent is guided by intrinsic rewards based on novelty or surprise, rather than by predefined random exploration strategies.
Lifelong Curiosity
Ability of an agent to maintain exploratory motivation over long periods, continuously adapting its behaviors to discover new knowledge in changing environments.
Novelty Detection
Process of identifying observations or states significantly different from past experiences, serving as a basis for generating curiosity signals.
Go-Explore
Exploration algorithm that explicitly memorizes visited states with their corresponding trajectories, then systematically explores from these anchor points to discover new regions.