agent_run

class agentos.agent_run.AgentRun(run_type: str, parent_run: Optional[str] = None, agent_name: Optional[str] = None, environment_name: Optional[str] = None)

An AgentRun provides an API that agents can use to log agent related data/stats/tags/etc. AgentRun can be one of two flavors (which we call run_type), ‘evaluate’ and ‘learn’.

The AgentRun can contain tags that reference other AgentRuns for tracking the training history of an agent.

An AgentRun inherits from Run, and adds functionality specific to runs of agents, such as runs that evaluate the agent’s performance in an environment, or runs that cause the agent to learn in an environment.

Like a Run, an AgentRun can be used as a context manager, so that the developer doesn’t need to remember to mark a run as finished, for example:

with AgentRun('evaluate',
              parent_run=self.__component__.active_run) as run:
     # run an episode
     run.log_episode(
           # episode_data
           ...
     )

AGENT_NAME_KEY = 'agent_name'

ENV_NAME_KEY = 'environment_name'

EVALUATE_KEY = 'evaluate'

IS_AGENT_RUN_TAG = 'pcs.is_agent_run'

LEARN_KEY = 'learn'

RESET_KEY = 'reset'

RESTORE_KEY = 'restore'

RUN_TYPE_TAG = 'run_type'

__init__(run_type: str, parent_run: Optional[str] = None, agent_name: Optional[str] = None, environment_name: Optional[str] = None) → None

Create a new AgentRun.

Parameters

• run_type – must be ‘evaluate’ or ‘learn’

• parent_run – Optionally, specify the identifier of another Run that this run is a sub-run of. Setting this will result in this AgentRun being visually nested under the parent_run in the MLflow UI.

• agent_name – The name of the agent component being evaluated or trained. Defaults to “agent”.

• environment_name – The name of the environment component being evaluated or trained. Defaults to “environment”.

add_episode_data(steps: int, reward: float)

end(status: str = 'FINISHED', print_results: bool = True) → None

This is copied and adapted from MLflow’s fluent api mlflow.end_run

get_training_info() -> (<class 'int'>, <class 'int'>)

log_agent_name(agent_name: str) → None

log_environment_name(environment_name: str) → None

log_run_metrics()

log_run_type(run_type: str) → None

print_results()

class agentos.agent_run.RunStats(episode_count, step_count, max_reward, median_reward, mean_reward, min_reward, training_episode_count, training_step_count)

episode_count

Alias for field number 0

max_reward

Alias for field number 2

mean_reward

Alias for field number 4

median_reward

Alias for field number 3

min_reward

Alias for field number 5

step_count

Alias for field number 1

training_episode_count

Alias for field number 6

training_step_count

Alias for field number 7