Curriculum & warm start

Shock curriculum

Training with full-magnitude shocks from step 0 often diverges. The shock curriculum ramps shock magnitude from a small fraction up to 1.0 over the first N episodes:

curriculum_episodes: 200
curriculum_start: 0.1     # start at 10% of full magnitude

After episode curriculum_episodes, shocks are at full scale.

Warm start

Two flavours:

warm_start: true

Fits the network to the steady-state policy via L-BFGS in ~10-50 steps. Effectively initializes the network to a constant function.

warm_start: true

For network.type: linear_plus_mlp, warm start is automatically skipped — the residual architecture already starts at the linear policy by construction.

warm_start_dynare: <path>

Imports a Dynare-solved linear policy and fits the network to it. For research workflows where Dynare is the ground truth.

SS reset fraction

ss_reset_frac: 0.15

A fraction of episode rollouts re-initialize from a noisy steady state instead of continuing from where the previous episode ended. Prevents trajectories from drifting permanently outside the training support.