Global optimization for supply chain process by deep reinforcement learning

Abstract

The bullwhip effect is known as one of the problems in the supply chain. As a result of demand forecasting and decision-making, demand propagates from downstream to upstream while amplifying. This phenomenon is well reproduced by the Beer Game invented in the 1960’s. On the other hand, in online shopping, there is a gap between the information-flow in cyberspace and the object-flow in physical space. This gap can be a factor to promote the bullwhip effect , but it is difficult to reproduced with the original Beer Game. Therefore, we set up the new game called “Netshop Game” which extended the rules and the environment. On the new game, by using deep reinforcement learning, we are able to reproduce the local optimum that can occur in net shopping supply chain, and confirmed that it is effective for discovering a global optimum by introducing a meta viewpoint.