FEDERATED Q-LEARNING WITH REFERENCE-ADVANTAGE DECOMPOSITION: ALMOST OPTIMAL REGRET AND LOGARITHMIC COMMUNICATION COST

Keyphrases

• Regret 100%
• Communication Cost 100%
• Almost Optimal 100%
• Optimal Regret 100%
• Federated Q-learning 100%
• Information Bound 66%
• Q-learning Algorithm 66%
• Recent Advances 33%
• Variance Reduction 33%
• Novel Design 33%
• Optimal Policy 33%
• Existing Algorithms 33%
• Markov Decision Process 33%
• Multiple Agents 33%
• Central Server 33%
• Heterogeneous Communication 33%
• Forced Synchronization 33%
• Communication Switching 33%
• Policy Switching 33%
• Event-triggered Communication 33%
• Triggering Conditions 33%

Computer Science

• Communication Cost 100%
• Learning Algorithm 66%
• Reinforcement Learning 33%
• Variance Reduction 33%
• Markov Decision Process 33%
• Multiple Agents 33%
• Synchronism 33%
• Event-triggered 33%