Adaptive workload adjustment for cyber-physical systems using deep reinforcement learning

Abstract Reducing computational energy consumption in cyber-physical systems (CPSs) has attracted considerable attention in recent years. Associated with energy consumption is a heating of the devices. Device failure rate increases exponentially with increase temperature, so that high energy consumption leads to a significant shortening of processor lifetime. Reducing thermal stress without harming application safety and performance is the goal of this work. Our approach is to abort control tasks dispatch when this is judged, by a neural network, to not contribute to either safety or performance. This technique is orthogonal to others that have been used to reduce energy consumption such as dynamic voltage/frequency scaling and adaptive use of redundancy. Simulation experiments show that this approach leads to a further reduction in device aging when used in conjunction with these prior techniques.