multicoreRuns

statrl.experiments.parallelruns.multicoreRuns(env, learner, interact, nbReplicates, timeHorizon, oneRunFunction, root_folder)

Run one agent for many independent replicates, spread across CPU cores.

Each replicate gets its own deep copy of the environment, agent, and interaction.

Parameters:

• env (object) – Environment to replicate.

• learner (object) – Agent to replicate.

• interact (statrl.experiments.onerun.Interaction) – Interaction loop of the setting.

• nbReplicates (int) – Number of independent runs.

• timeHorizon (int) – Number of rounds per run.

• oneRunFunction (callable) – Function executing one replicate, called as oneRunFunction(env, learner, interact, timeHorizon, root_folder). In practice oneRunWithDump().

• root_folder (str) – Directory the per-replicate dumps are written to.

Returns:

• scores (list of str) – One dump filename per replicate, in the order the jobs were created.

• elapsed (float) – Mean wall-clock seconds per replicate. Since the runs are concurrent this is total elapsed time divided by nbReplicates, so it measures throughput rather than the cost of a single run.

Notes

Uses all available cores (n_jobs=-1). Everything passed in must be picklable, which is why BatchMAB accepts a plain list of batch sizes rather than only a callable.