Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Distribution over Actions
A probabilistic representation of the set of all possible actions in a given state, allowing the agent to choose among several valid behaviors instead of a single deterministic action.
Gaussian Mixture over Actions
A modeling technique where the action distribution is represented by a mixture of several Gaussian distributions, enabling the capture of multiple and complex behavior modes.
Importance Resampling
A statistical technique for estimating the properties of one distribution using samples from another distribution, applied to correct bias when learning from multi-modal demonstrations.
Divergence Loss Function
An objective function that minimizes the divergence between the distribution of actions predicted by the model and the distribution of actions observed in the demonstrations, favoring the learning of multi-modal behaviors.
Mode Collapse
A phenomenon where a multi-modal learning model forgets certain modes of the distribution and focuses on a subset of behaviors, thereby reducing the diversity of generated actions.
Conditional Neural Network
A neural network architecture whose parameters or outputs are conditioned on input variables, used to model action distributions that depend on the state of the environment.
Heterogeneous Demonstrations
An imitation learning dataset containing demonstrations from different strategies or experts, naturally introducing multi-modality into the behaviors to be learned.
Trajectory Clustering
A technique for grouping sequences of actions and states to identify the different behavior modes present in the demonstrations, facilitating the learning of multi-modal policies.
Mixture Flow Neural Network
An architecture that combines multiple flow networks to model a complex distribution, particularly suitable for representing multi-modal action distributions in imitation learning.