DDPG with Transfer Learning and Meta Learning Framework for Resource Allocation in Underlay Cognitive Radio Network

Cognitive Radio (CR) is an intelligent device equipped with a Cognitive Engine (CE) capable of making decisions and finding the best policy for a dynamic network. Superior decision-making policy takes extensive learning time. A suitable learning algorithm reduces the learning time and provides a boost to CE capabilities. The underlay CR model allows PU and SU to coexist in the same frequency band by restricting SU interference below an acceptable level. This paper presents an underlay Resource Allocation (RA) model that employs Transfer Learning (TL) and Meta Reinforcement Learning (MRL) to solve a non-convex optimization problem. The allocation of resources is performed by incorporating TL and MRL into the existing Deep Deterministic Policy Gradient (DDPG) method. The merging of TL and MRL accelerates the network’s learning process and allows it to adapt rapidly to the changing environments. The proposed algorithms are compared to basic Q learning, dueling Deep Q Networks, and hybrid algorithms in terms of Quality of Experience (QoE) metric, learning speed, congestion rate, and stability. The simulation findings indicate that our proposed approach outperforms the existing techniques. The adaptability of the network is also tested by changing the environment from Additive White Gaussian Noise (AWGN) to Rayleigh fading. In addition, trade-offs between network scalability, congestion, and performance are evaluated.