Saving and Loading

A trained policy is just its cuts, so there is no need to retrain every session. save() writes a trained instance to a single file, and SDDP.load() brings it back.

# after training
sddp.save("clearlake.sddp")

# later, in a fresh session
from gamspy.formulations import SDDP

sddp = SDDP.load("clearlake.sddp")
sddp.policy(stage="mar", state=150, noise=350, report=[R, L, Z, F])

The path must end in .sddp. The file bundles the host container (with all the trained cuts) and a small sidecar recording which symbols play which role, so a reloaded instance is ready to use immediately.

Read-only instances

A loaded instance is read-only: policy() and simulate() work, but add_state(), set_noise(), build() and train() raise. To change the model or add iterations, rebuild and retrain from scratch.

Note

The file records the module version it was written with. A file from an incompatible (older major) version is rejected with a clear message rather than failing obscurely; retrain to produce a current file.

See also

Using a Trained Policy covers the queries a loaded policy supports; the ClearLake tutorial saves and reloads a policy end to end.