Application of Federated Reinforcement Learning for Heterogeneous Data Distribution in Distributed Team Efficiency Optimization

The challenges in the Federated Reinforcement Learning (FRL) of heterogeneous data distribution include uneven data distribution, poor performance of model generalization, and high communication overhead in optimizing the efficiency of multi-node distributed teams. In response to these concerns, this paper suggests a federated reinforcement learning implementation on the basis of multi-layer policy aggregation. This approach balances between personalization and global optimization in four steps including independent training of local agents, gradient adaptive weighted aggregation, cross-node policy distillation, and adjusting communication frequency dynamically. Each node, to begin with, learns a local model through Deep Deterministic Policy Gradient (DDPG) depending on the local task nature. Second, gradient similarity computation is used to achieve the adaptive weight allocation in the conditional of heterogeneous data. Following this, a policy distillation mechanism is applied to project, in a common global policy, multi-node policies and maximize convergence. Lastly, it is dynamic communication scheduling that lowers the bandwidth usage. The method is proven to be effective in six distributed team task scenarios, which are confirmed by experiments. The global reward is better, the average convergence cycle of the algorithm is shortened by about 93.5 rounds, and the team task completion rate range becomes shorter by 0.9% when compared to the standard FRL algorithm. The findings reveal that this approach is capable of accommodating heterogeneous data distribution and enhancing the collaborative performance of distributed workforce to a significant extent.