Resource Allocation in OFDMA Networks with Deep Reinforcement Learning

In a distributed OFDMA network, wireless channel is divided into multiple time-frequency resource block for transmission. Because there is no information interaction between nodes in a distributed network, and independent resource allocation of each nodes will result in collisions. Reasonable allocation resources can effectively reduce the collision ratio and improve system throughput. For the distributed network, we propose a time-frequency resource allocation method based on deep reinforcement learning (DRL), with definition of the state space, action space, and rewards. Each node uses this algorithm to realize time-frequency resource allocation. The nodes in the network use this algorithm without information interaction between the other nodes. Different nodes have a preference for the occupation of resource blocks, which reduces the probability of collisions in resource allocation. When the traffic intensity fluctuates, we introduce the concept of virtual traffic, which improves the continuity of the node’s selection of resource blocks. This continuity reduces the impact of traffic fluctuations on the node’s decisions. Compared with the random access slotted-aloha protocol, simulation results show the network throughput has been significantly improved. When the traffic intensity remains constant or fluctuates, the throughput can approximate that of the centralized network.