Running experiments#

Module: statrl.experiments

The experiments package is the harness for benchmarking agents: it runs an algorithm for many independent replicates (in parallel), measures regret against an oracle, and plots the result. Everything is glued together by one entry point.

The orchestrator#

runLargeMulticoreExperiment():

runLargeMulticoreExperiment(
    env, agents, oracle, interact,
    timeHorizon=1000, nbReplicates=100, root_folder="results/",
)

For each agent it launches nbReplicates runs, runs the oracle for the same horizon, computes the regret (oracle score minus agent score), writes a logfile, and saves regret plots under root_folder.

The pipeline underneath#

Function	Role
`multicoreRuns()`	Runs `nbReplicates` replicates of one learner across CPU cores with `joblib`, building environments via `gymnasium.make`. Returns the scores and the mean elapsed time.
`oneRunWithDump()`	Runs a single interaction and pickles the score series to disk (one replicate).
`computeScoreDiffs()`	Loads the pickled scores and computes regret statistics vs. the oracle: mean, median, and the 0.25 / 0.75 quantiles over time.
`plotScoreDiffs()`	Produces the regret plots (linear and log-scaled `y`) as PNG/PDF with `matplotlib`.
`dump()`	Pickle helper used to persist intermediate score series (with `clear_auxiliaryfiles()` to clean them up).

Results are written under root_folder (default "results/"): the intermediate per-replicate dumps, a logfile, and the regret figures.