Multi-agent reinforcement learning for intelligent resource allocation in IIoT networks

In the industrial Internet of Things (IIoT), a high number of devices with limited resources, like computational power, memory, bandwidth and, in case of wireless sensor networks, also energy, communicate. At the same time, the amount of data as well as the demand for data processing in the edge is rapidly increasing. To enable Industry 4.0 (I4.0) and the IIoT, an intelligent resource allocation is required to make optimal use of the available resources. For this purpose, a multi-agent system (MAS) based on deep reinforcement learning (DRL) is proposed. Multi-agent reinforcement learning (MARL) is already taken into account in different communication networks, e.g. for intelligent routing. Despite its great potential, little attention is paid to these methods in industry so far. In this work, DRL is applied for resource allocation and load balancing for industrial edge computing. An optimal usage of the available resources of the IIoT devices should be achieved. Due to the structure of IIoT systems as well as for security reasons, a MAS is preferred for decentralized decision making. In subsequent steps, it is planned to add and remove devices during runtime, to change the number of tasks to be executed as well as evaluations on single- and multipolicy-approaches. The following aspects will be considered for evaluation: (1) improvement of the resource usage of the devices and (2) overhead due to the MAS.