Deep Reinforcement Learning for Resource Management in C-RAN for Dynamic Environment

The cloud radio access network (C-RAN) is a novel mobile network architecture that reduces network design and operation costs and makes it easier to deploy large-scale networks. However, as network scale gets larger, its surrounding environment, such as user traffic demand or user mobility, will become more dynamic and diverse. Because a larger amount of over-provisioning is needed to prepare in advance for such changes, network management methods that can flexibly follow the varying environment are becoming important. With the great attention given to artificial neural network techniques, some researchers have attempted to apply neural networks to the management of C-RANs. However, neural network have difficulties with training in dynamically varying situations because they typically loses previously acquired skills when learning a new skill. Recently, a study found that evolution under an environment with modularly varying goal (MVG) which share the same subgoals while having different combination of them leads to a neural network configuration that is able to quickly relearn multiple goals, and also exhibit high adaptability to new goals. In this research, we apply MVG to neural network management of a C-RAN to enable flexible management following the dynamic varying environment. Specifically, we consider the power consumption problem solved by deep reinforcement learning, and propose a design approach that is able to obtain such network configurations. From the evaluation, we show that the MVG-based approach is able to quickly relearn different goals, and can have a smaller reward decrease when facing a new environment.