Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Approximated Behavioral Metric-based State Projection for Federated Reinforcement Learning

Zengxia Guo, Bohui An, Zhongqi Lu

TL;DR

The paper tackles the challenge of privacy-preserving federated reinforcement learning in heterogeneous environments by proposing FedRAG, which shares parameters of a behavior-metric based state projection function rather than raw data. FedRAG leverages a RAG-based representation learning objective to produce task-relevant embeddings, while central aggregation enforces global alignment through a regularization term. The approach achieves superior cross-environment generalization on DeepMind Control Suite tasks, maintains in-environment performance, and demonstrates resilience to increasing client heterogeneity. The work combines theoretical insights on the RAG distance with practical federated training and privacy considerations, offering a scalable, privacy-conscious pathway for collaborative RL across diverse domains.

Abstract

Federated reinforcement learning (FRL) methods usually share the encrypted local state or policy information and help each client to learn from others while preserving everyone's privacy. In this work, we propose that sharing the approximated behavior metric-based state projection function is a promising way to enhance the performance of FRL and concurrently provides an effective protection of sensitive information. We introduce FedRAG, a FRL framework to learn a computationally practical projection function of states for each client and aggregating the parameters of projection functions at a central server. The FedRAG approach shares no sensitive task-specific information, yet provides information gain for each client. We conduct extensive experiments on the DeepMind Control Suite to demonstrate insightful results.

Approximated Behavioral Metric-based State Projection for Federated Reinforcement Learning

TL;DR

The paper tackles the challenge of privacy-preserving federated reinforcement learning in heterogeneous environments by proposing FedRAG, which shares parameters of a behavior-metric based state projection function rather than raw data. FedRAG leverages a RAG-based representation learning objective to produce task-relevant embeddings, while central aggregation enforces global alignment through a regularization term. The approach achieves superior cross-environment generalization on DeepMind Control Suite tasks, maintains in-environment performance, and demonstrates resilience to increasing client heterogeneity. The work combines theoretical insights on the RAG distance with practical federated training and privacy considerations, offering a scalable, privacy-conscious pathway for collaborative RL across diverse domains.

Abstract

Federated reinforcement learning (FRL) methods usually share the encrypted local state or policy information and help each client to learn from others while preserving everyone's privacy. In this work, we propose that sharing the approximated behavior metric-based state projection function is a promising way to enhance the performance of FRL and concurrently provides an effective protection of sensitive information. We introduce FedRAG, a FRL framework to learn a computationally practical projection function of states for each client and aggregating the parameters of projection functions at a central server. The FedRAG approach shares no sensitive task-specific information, yet provides information gain for each client. We conduct extensive experiments on the DeepMind Control Suite to demonstrate insightful results.
Paper Structure (27 sections, 2 theorems, 18 equations, 11 figures, 1 table, 1 algorithm)

This paper contains 27 sections, 2 theorems, 18 equations, 11 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Learning.
  4. Federated Representation Learning.
  5. Federated Reinforcement Learning.
  6. Behavioral Metrics-based Representation Learning.
  7. Preliminaries
  8. Methodology
  9. Problem Formulation
  10. Client RAG Distance
  11. FedRAG Framework
  12. Effectiveness of Anti-attack
  13. Experiment
  14. Experimental Settings
  15. FedRAG vs. Baseline Performance Comparison
  16. ...and 12 more sections

Key Result

Theorem 1

The function $d^\pi$ is a contraction mapping with respect to the $L_\infty$ norm and has a unique fixed-point $D^\pi$.

Figures (11)

  • Figure 1: Framework of FedRAG. Periodically, the local state projection function parameters $\omega^k$ are synchronized to a central server. Then the central server distributes the averaged parameters to the clients. For each client, a regularization term is incorporated to ensure that the client's local state projection parameters follow the global updates.
  • Figure 2: Comparison of FedRAG with Baseline in other environments.
  • Figure 3: The results of varying lambda. In the above experiment, the training data and testing data are from environments with same setting, while in the below experiment, they are come from environments with different settings.
  • Figure 4: Comparison of Local and FedRAG with $\lambda=0.1/0.001$ in same or other environments.
  • Figure 5: Experimental results on various DMC tasks.
  • ...and 6 more figures

Theorems & Definitions (3)

  • Theorem 1
  • proof
  • Theorem 2: Value function difference bound