Implementation and evaluation of NoisyNets to reinforcement learning of automated ICT system design

Abstract

This paper proposes to apply the method with an additional noisy layer to the structure of the graph neural network for reinforcement learning in automated design technology for information and communication systems. The automated design technology has an elementary problem of huge learning time caused by overestimating a specific configuration because of hardly ever rewards despite huge exploration space with a vast combination of selections, arrangements, and connections. The parametric noise applied to the network is learned with gradient descent and the remaining network weights to reduce this harmful overestimation during learning and increase the design exploration efficiency. The evaluation result showed that using the proposed algorithm for our automated design technology in development could shorten 15% of the episodes needed for learning to converge.