AI-ordlista
Den kompletta ordlistan över AI
Inverse Reinforcement Learning
Method that consists of deducing the reward function of an expert from its optimal trajectories, then allowing the agent to learn an optimal policy.
State-only Imitation Learning
Learning paradigm where the agent only has access to the states visited by the expert without knowledge of the actions taken, requiring specific approaches to infer behaviors.
Trajectory Matching
Approach that minimizes the divergence between the trajectory distributions generated by the agent and those of the expert, often used in learning without access to actions.
GAIL
Framework combining imitation learning and generative adversarial networks, where a discriminator distinguishes the trajectories of the expert from those of the agent.
Dataset Aggregation
Iterative algorithm that collects new expert data based on the errors of the current agent, progressively aggregating a more robust dataset.
Forward-Forward Algorithm
Unsupervised learning method that predicts future states from current states without requiring action data, used in imitation by observation.
Observation-based Learning
Learning process where the agent acquires skills by observing only environmental states and results, without direct access to the expert's actions.
State Distribution Matching
Technique aiming to align the distribution of states visited by the agent with that of the expert, used when actions are not observable.
No-action Imitation
A form of imitation learning where the agent must learn to reproduce expert behavior without any information about the actions taken.
Passive Learning
Learning mode where the agent passively observes demonstrations without active interaction with the environment, typical of imitation by observation.
Expert Demonstration
Set of trajectories or states provided by an expert serving as reference for imitation learning, crucial in approaches without access to actions.
State-Action Distribution
Joint distribution of states and actions that the agent seeks to approximate, often inferred from the state distribution alone in imitation by observation.
Trajectory-based Learning
Learning approach that focuses on reproducing complete trajectories rather than individual state-action decisions, adapted to observation without actions.
Dynamics Model
Model learning the transition between consecutive states in expert demonstrations, essential for inferring actions when they are not observed.
Occupancy Measure
Statistical measure quantifying the visitation frequency of each state-action, adapted to contexts where only state visitations are observable.