Adaptive reinforcement learning method for networks-on-chip

In this paper, we propose a congestion-aware routing algorithm based on Dual Reinforcement Q-routing. In this method, local and global congestion information of the network is provided for each router, utilizing learning packets. This information should be dynamically updated according to the changing traffic conditions in the network. For this purpose, a congestion detection method is presented to measure the average of free buffer slots in a specific time interval. This value is compared with maximum and minimum threshold values and based on the comparison result, the learning rate is updated. If the learning rate is a large value, it means the network gets congested and global information is more emphasized than local information. In contrast, local information is more important than global when a router receives few packets in a time interval. Experimental results for different traffic patterns and network loads show that the proposed method improves the network performance compared with the standard Q-routing, DRQ-routing, and Dynamic XY-routing algorithms.