Robustness and stress testing

A policy that looks good on the environment it was tuned for can fail once the world shifts. stress_test evaluates a policy across a set of perturbed environments and reports the mean return on each, so a robust policy and a brittle one can be told apart.

from decisionrl.evaluation import stress_test
from decisionrl.envs import NonstationaryInventory
from decisionrl import baselines as B

variants = {
    "nominal":      NonstationaryInventory,
    "demand_spike": lambda: NonstationaryInventory(demand_low=8.0, demand_high=24.0),
    "volatile":     lambda: NonstationaryInventory(switch_prob=0.20),
}
returns = stress_test(B.tracking_base_stock(3.0), variants, episodes=30)

Pass any (env, obs) -> action policy. To stress-test a trained agent, wrap it: lambda env, obs: agent.predict(obs).

Example: fixed vs adaptive inventory policy

Mean return under three perturbations (the fixed base-stock is tuned on the nominal setting):

Variant	Fixed base-stock	Adaptive tracking
nominal	268.9	322.4
demand spike (higher demand)	342.6	476.9
volatile (fast regime switching)	238.9	231.4

The adaptive policy is far more robust to a demand-level shift, which is exactly what it is built for: it reads recent demand and follows it. Under very fast regime switching it is marginally worse, because the smoothed demand estimate lags the rapid changes. That trade-off is the point of a stress test: it surfaces where a policy holds up and where it does not, instead of reporting a single number from the setting it was tuned on.