Utilizing Reinforcement Learning Techniques for Supply Chain with Amazon SageMaker
summary
Claim: Amazon SageMaker RL runs Ray RLlib and TensorFlow/PyTorch in managed containers, uses managed spot instances with checkpoint resume, and supports homogeneous/heterogeneous scaling. A simulator is built from two years of retailer logs by historical-to-simulator via webpage-view correction to recover true demand, then RL agents (A3C, SAC, ARS) train on that simulator. Trained policies (two years train, 19 weeks test) are competitive with or exceed a newsvendor baseline and A3C and SAC outperform newsvendor in some tests, showing the approach works in simulation.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:48Amazon SageMaker overview
- 4:41Cost and infrastructure benefits
- 6:48Reinforcement learning on SageMaker
- 11:03Inventory management problem
- 15:53Markov decision process formulation
- 22:55Historical data to simulator
- 26:45Experiment policies and results
- 30:35Recap and conclusions
related talks
25mIntroducing Amazon SageMaker Kubeflow Reinforcement Learning Pipelines for Robotics
19mUsing Amazon SageMaker to develop and deploy ML models
30mProductionizing R workloads using Amazon SageMaker, featuring Siemens
32mAccelerating autonomous vehicle solutions with SageMaker, featuring Lyft
39mTrain ML models at scale with Amazon SageMaker Training
33mBuild, train, and deploy ML models using Amazon SageMaker
19mDeveloping large-scale machine learning algorithms on Amazon SageMaker
29mGet rolling with ML on AWS DeepRacer