Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
One-shot Imitation Learning
Learning paradigm where an agent acquires the ability to execute a new task after observing a single demonstration, without requiring additional task-specific training.
Contextual Policy
Conditional policy function that takes as input both the current state of the environment and the context of the observed demonstration to generate appropriate actions.
Demonstration Encoding
Process of transforming a demonstration sequence into a semantic representation that can be used by the model to guide the execution of the imitated task.
Behavior Cloning from Demonstration
Supervised learning technique where the model directly learns to map states to actions by imitating expert behavior from demonstration examples.
Episode-based Learning
Training method where tasks are presented as complete episodes, including demonstration and execution phases, to facilitate one-shot imitation learning.
Cross-domain Imitation
Ability to imitate tasks even when demonstrations come from slightly different domains or present significant contextual variations.
Trajectory Alignment
Process of spatial and temporal adjustment between the demonstration trajectory and the execution trajectory to ensure accurate imitation despite initial variations.
Latent Task Representation
Latent space where tasks are encoded in an abstract manner, allowing the capture of structural invariants and generalization to new task instances.
Zero-shot Generalization
Extension of one-shot learning where the model can perform on never-seen tasks even without demonstration, relying on learned structural similarities.
Conditional Neural Process
Stochastic neural architecture capable of modeling distributions over functions, particularly suitable for learning from few examples in imitation.
Hierarchical Imitation
Multi-level learning structure where complex tasks are decomposed into simpler sub-tasks, facilitating imitation from single demonstrations.