run_drl

class xuance.engine.run_drl.RunnerDRL(config: Namespace, envs: DummyVecEnv | SubprocVecEnv | None = None, agent=None, manage_resources: bool = None)[source]

Bases: RunnerBase

Runner for single-agent Deep Reinforcement Learning (DRL).

RunnerDRL orchestrates the full lifecycle of an experiment, including environment creation, agent initialization, training, testing, and benchmarking. It is responsible for experiment-level logic rather than algorithmic details.

Responsibilities:

  • Create and manage environments and agent (unless injected externally).

  • Control training, testing, and benchmarking workflows.

  • Handle experiment-level logic (run mode, evaluation loop, model saving).

  • Manage resource lifecycle (envs.close(), agent.finish()) based on ownership semantics.

Notes

  • Algorithm-specific logic should remain inside the Agent.

  • Runner focuses on experiment orchestration and reproducibility.