Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Federated Reinforcement Learning
Machine learning paradigm combining reinforcement learning principles with federated learning, allowing multiple agents to learn collectively without sharing their raw data. Agents perform local learning then aggregate their knowledge to improve a common global policy.
Federated Agent
Autonomous learning entity operating in a distributed environment that participates in a federated learning process while preserving the confidentiality of its local data. The agent makes decisions based on its local policy while contributing to the improvement of the global model.
Global Policy
Consolidated decision strategy resulting from the aggregation of local policies from different participating agents in a federated learning system. This policy represents collective knowledge and is periodically distributed to agents to guide their future actions.
Policy Aggregation
Algorithmic process of combining policies or learning parameters from multiple local agents to form an improved global policy. Aggregation typically uses weighted averaging techniques or more sophisticated methods like FedAvg adapted for reinforcement learning.
Synchronous Communication
Coordination mode where all federated agents must complete their local learning cycle before proceeding to global aggregation. This approach ensures temporal consistency but can introduce bottlenecks due to slower agents.
Differential Privacy
Formal framework ensuring that an individual agent's contribution cannot be significantly inferred from the aggregated global model. In the federated context, it protects against inference attacks while enabling effective collaborative learning.
Federated Multi-Agent Learning
Extension of federated reinforcement learning where multiple agents interact in potentially different environments but share knowledge to collectively improve their performance. This approach combines multi-agent coordination challenges with federated privacy constraints.
Federated Convergence
Property ensuring that the federated reinforcement learning algorithm reaches an optimal or near-optimal policy despite data distribution and limited communication. Convergence depends on factors like data heterogeneity, communication frequency, and aggregation method.
Federated Gradient
Gradient of the global objective function computed in a distributed manner from the local gradients of the various participating agents. Federated gradients allow updating the model parameters without exposing the agents' sensitive data.
Local Experience
Set of state-action-reward transitions accumulated by an individual agent in its specific environment during a local learning cycle. This experience remains private and is used solely to compute local updates before aggregation.
Decentralized Coordination
Mechanism allowing agents to align on common goals without explicit centralization, using peer-to-peer communication protocols or consensus. In federated learning, it emerges through the iterative aggregation of local policies.
Weighted Aggregation
Technique for aggregating local contributions where each agent is assigned a weight based on the quality or quantity of their data, their performance, or other relevant metrics. This approach gives more influence to more reliable or representative agents.
Federated Exploration
Distributed exploration strategy where agents explore different state-action spaces in a coordinated manner to maximize collective discovery while minimizing redundancy. Federated exploration optimizes learning efficiency in complex and distributed environments.
Distributed Reward
Reward system where agents receive local feedback based on their actions while contributing to a shared global objective. Distributed rewards must balance individual incentives with the collective performance of the federated system.
Federated Knowledge Transfer
Process of selectively transferring knowledge learned by one agent or group of agents to other agents in the federated network. This transfer optimizes learning efficiency by capitalizing on successful experiences while respecting privacy constraints.
Failure Robustness
Ability of the federated learning system to maintain its performance despite disconnections, malicious behaviors, or degradations of some participating agents. Robustness is essential to ensure reliability in uncontrolled distributed environments.