Performance Tuning¶
This guide focuses on practical knobs that reduce runtime and memory.
FQE training knobs¶
FQEConfig exposes the main levers:
hidden_sizes: network width and depthbatch_size: lower to fit in memorynum_epochs: reduce for faster iterationsnum_iterations: fewer fitted-Q iterationsweight_decay: regularization for stabilityseed: reproducibility
Example:
from crl.estimators.fqe import FQEConfig, FQEEstimator
config = FQEConfig(
hidden_sizes=(32, 32),
batch_size=64,
num_epochs=5,
num_iterations=5,
)
estimator = FQEEstimator(estimand, config=config, device="cpu")
Diagnostics cost¶
Diagnostics can add overhead. If you only need estimates:
Bootstrap CI cost¶
Bootstrap can be expensive. Reduce the number of resamples: