maxvolk / sddp

Seminar thesis on incorporating experience replay into stochastic dual dynamic programming

Stochastic Dual Dynamic Programming

This repository is a work in progress and aims at an implementation of Stochastic Dual Dynamic Programming for Multi-Stage Stochastic Linear Programs. In particular the extension including batch learning via incorporation of experience replay will be implemented.

TASK DESCRIPTION: Stochastic dual dynamic programming (SDDP) is a widely employed algorithm to solve multistage stochastic programs. While it originated in stochastic programming, it can also be interpreted as a type of Q-learning algorithm in reinforcement learning. This means that batch learning ideas from this field can be incorporated into SDDP as well. This seminar topic provides an overview on how this incorporation works and how it improves classical SDDP.