Using a Trained Policy#
Training produces a policy: the collection of cuts. Two methods put it to work:
policy() answers a single “what should I do here?” question, and
simulate() evaluates the policy across many random paths.
Point queries with policy()#
policy() solves one stage at a given situation: a stage, the incoming state,
and the realised noise.
decision = sddp.policy(stage="mar", state=150, noise=350, report=[R, L, Z, F])
decision.decisions # {'R': 200.0, 'L': 250.0, 'Z': 0.0, 'F': 50.0}
decision.approx_cost_to_go # 875.0
The returned PolicyResult carries the queried stage, the
incoming_state and noise, the optimal decisions for the reported
variables, and approx_cost_to_go, the immediate stage cost plus the
cut-approximated future cost from here on.
stateis a scalar for a single state variable, or adictkeyed by state-variable name when there are several.reportlists the variables whose optimal level to return; it defaults to the state variables. A time-only variable reports afloat; a variable with extra dimensions reports adictkeyed by the non-time labels.
Note
policy() needs a trained policy. Calling it before train() warns that
no cuts exist yet and returns a decision that is not the trained policy.
Monte Carlo with simulate()#
simulate() runs the policy forward on fresh sampled paths and returns the
realised cost distribution.
sim = sddp.simulate(n_paths=1000, seed=0)
print(sim.summary)
n_paths 1000.000000
mean 140.000000
std 438.283119
p5 0.000000
p50 0.000000
p95 1500.000000
max 2500.000000
Name: total_cost, dtype: float64
Each simulated path is one sequence of sampled noise values across all stages.
simulate() returns a SimulationResult containing:
total_cost: the total cost of each path, summed across all stages.stage_costs: the cost at each stage of each path.noise: the sampled noise value at each stage of each path.variables: one table for each reported variable, containing its value at the end of each stage for each path.
sim.summary summarises the total costs across paths. It reports the number
of paths, mean, standard deviation, percentiles, and maximum cost.
Use report to choose which variables to record. By default, these are the
registered state variables. If seed is omitted, simulation uses the seed
configured when creating the SDDP instance plus one; 42 -> 43.
After loading a saved policy with SDDP.load(...), you can call policy()
and simulate() on the loaded instance without retraining. See
Saving and Loading for details.
policy() and simulate() apply the
solve status checks to every required solve.
If a check fails, the method raises ValidationError instead of
returning a result.
A solver-reported user interruption (SolveStatus.UserInterrupt)
also raises ValidationError for these methods.
See also
The ClearLake tutorial runs both methods on a complete
model; Risk Aversion (CVaR) interprets the skewed cost distribution that
simulate() reveals here.