GAMSPy and Stochastic Programming

Beta feature

SDDP has been added in GAMSPy 1.25.0 and is a beta feature under active development. Its public API may still change between releases without preserving backward compatibility, so pin your GAMSPy version for reproducible work and validate results before relying on them for real decisions. Feedback is very welcome on the GAMS forum.

Many decisions have to be made over time and under uncertainty: how much water to release from a reservoir before the season’s rainfall is known, how much to store before tomorrow’s demand, how to commit resources as the future unfolds. GAMSPy supports these multistage stochastic problems through SDDP (Stochastic Dual Dynamic Programming).

You write the single-stage problem as an ordinary GAMSPy model, describe how the state carries forward and how the uncertainty is distributed, and SDDP learns a policy: a rule that, for any state and any realised outcome, returns the optimal decision and its expected cost from there on.

When to use SDDP

SDDP fits problems that

• unfold over several stages, with decisions made before each stage’s uncertainty is revealed;

• carry a state from one stage to the next (a reservoir level, an inventory, a stored quantity);

• have a convex (linear) stage problem and a finite set of noise scenarios.

If your problem is deterministic or has a single decision point, an ordinary GAMSPy model is the simpler choice. For a hands-on start, the ClearLake tutorial builds and trains a complete model in a few dozen lines.

• Introduction
• The SDDP Workflow
• State Variables
• Scenarios and Probabilities
• Risk Aversion (CVaR)
• Training and Convergence
• Using a Trained Policy
• Saving and Loading