Core Concepts
The authors propose TranSDDP, a novel Transformer-based stagewise decomposition algorithm, to efficiently generate piecewise linear approximations of the value function and solve large-scale multistage stochastic optimization problems. The TranSDDP model leverages the structural advantages of the Transformer architecture to integrate subgradient cutting planes, significantly reducing computation time while preserving solution quality.
Abstract
The authors focus on solving large-scale multistage stochastic programming (MSP) problems, which pose significant computational challenges due to the curse of dimensionality. They introduce TranSDDP, a novel Transformer-based stagewise decomposition algorithm, to address these challenges.
Key highlights:
Traditional stagewise decomposition algorithms, such as stochastic dual dynamic programming (SDDP), face growing time complexity as the subproblem size and problem count increase.
TranSDDP leverages the structural advantages of the Transformer model to implement a sequential method for integrating subgradient cutting planes to approximate the value function.
Through numerical experiments on energy planning, financial planning, and production planning problems, the authors demonstrate that TranSDDP efficiently generates piecewise linear approximations of the value function, significantly reducing computation time while preserving solution quality.
Compared to benchmark algorithms, TranSDDP and its variant TranSDDP-Decoder exhibit notable computational advantages, especially when solving a large number of similar problems with slight variations.
The authors also verify the feasibility of the cuts generated by the proposed models and provide a comparison of the value function approximations.
Stats
The number of variables and constraints for the numerical experiments are as follows:
For the 7-stage problems:
Energy Planning: 78,124 variables, 136,717 constraints
Financial Planning: 46,873 variables, 54,684 constraints
Production Planning: 128,904 variables, 121,092 constraints
For the 10-stage problems:
Energy Planning: 118,096 variables, 206,668 constraints
Financial Planning: 78,729 variables, 98,410 constraints
Production Planning: 206,667 variables, 186,985 constraints