C2C Return Logistics: a large Markov Decision Process analysis of an innovative concept
In this research, the innovative C2C Return Logistics (C2C) concept, in which returns are directly sent to other customers, is modelled and tested on various scenarios. The C2C concept shows promising results to serve as a solution to the return problem online retailers face, albeit dependent on product characteristics. The simulation model presented in this research can be applied to any online retailer for testing the effect of implementation of C2C. With the C2C concept modelled as a large, complex Markov Decision Process (MDP), two solution techniques from the field of Operations Research and Reinforcement Learning, being respectively Approximate Dynamic Programming (ADP) and Monte Carlo Tree Search (MCTS), are implemented and compared to each other in this research. It therefore contributes as reference work for large MDPs, by both testing the so-called `model-free' approach (MCTS), as well as the `structured-based' approach (ADP), from both fields on the same use case. Especially ADP shows promising results compared to the MCTS method, both in terms of performance and in terms of running times. If an MDP has a certain problem structure which makes any generalization of state space possible, ADP can serve as a powerful tool. The MCTS method, however, although applicable to any kind of MDP which consists of a simulator replicating state transitions, showed moderate performance in reasonable amount of running time. The performance improved, however, when the size of the state space of the problem decreased.
|Keywords||Return Logistics, Markov Decision Process, Approximate Dynamic Programming, Monte Carlo Tree Search|
|Thesis Advisor||Heuvel, W. van den|
Tetteroo, E.M. (2019, October 24). C2C Return Logistics: a large Markov Decision Process analysis of an innovative concept. Econometrie. Retrieved from http://hdl.handle.net/2105/50037