Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Offline Multi-Task Reinforcement Learning
Learning paradigm where multiple policies for different tasks are learned simultaneously from fixed batch datasets without interaction with the environment.
Batch Multi-Task Learning
Approach where the agent learns to solve multiple tasks using only pre-collected data, without online exploration during training.
Shared Dataset Policy Optimization
Multiple policy optimization technique using a common pool of experience data to improve learning efficiency across tasks.
Task-Agnostic Representation Learning
Process of learning state-action representations that are generalizable from batch data without specific knowledge of future tasks.
Conservative Multi-Task Policy Optimization
Method ensuring that multi-task policies do not deviate significantly from the behavior observed in the batch dataset to avoid out-of-support distributions.
Multi-Task Batch Constrained Q-Learning
Extension of BCQ to the multi-task context where the Q-function is constrained by batch data while sharing knowledge between tasks.
Multi-Task Distributional RL
Framework modeling the complete distribution of returns rather than their expectation for each task in an offline multi-task context.
Offline Multi-Task Meta-Learning
Learning of meta-knowledge from multi-task batch datasets to enable rapid adaptation to new tasks with few data points.
Task Decoupling
Technique separating task-specific representations from shared knowledge to optimize offline multi-task learning.
Multi-Task Offline Evaluation Metrics
Specific measures evaluating the performance of multi-task policies without interaction, such as multi-task FQE or weighted importance sampling.
Task-Specific Policy Heads
Network architecture with shared common trunk and distinct output heads for each task in offline multi-task learning.
Multi-Task Offline Data Efficiency
Measure of how efficiently batch data is used to learn multiple policies compared to single-task learning.
Cross-Task Knowledge Transfer
Process of automatically transferring useful knowledge between different tasks when learning from shared batch datasets.
Multi-Task Offline Value Function Factorization
Decomposition of the value function into shared and task-specific components to improve offline multi-task learning.
Task Clustering in Offline Settings
Automatic grouping of similar tasks based on their batch data to optimize knowledge sharing and resource allocation.
Multi-Task Offline Exploration-Exploitation
Dilemma adapted to the offline context where the balance between using existing data and controlled extrapolation is managed for multiple tasks.
Shared Dynamics Model
Single transition model learned from multi-task batch data capturing common and specific dynamics of environments.
Multi-Task Offline Curriculum Learning
Automatic sequencing of tasks during offline training based on their difficulty and interdependence to optimize learning.