Skip to main content

AgentInterface

class AgentInterface(ABC)

Abstract interface for the agent half of an RL optimization — everything that is not the Gymnasium environment step itself. Implement these methods on your RLAgentEnv subclass. See RL Agents for how RLExecutor drives your class through the optimization loop.

Model transfer, export, and import are not part of AgentInterface; use ModelHandler for those.

Import​

from adk.executors.rl import AgentInterface

Members​

MemberDescription
env_dataEnvData constructed by RLExecutor and passed at construction.
agent_dataAgentData constructed by RLExecutor and passed at construction.

Methods​

compute_action​

def compute_action(self, observation, info) -> action

Choose an action given the current observation and step info dict from the environment.

experience​

def experience(
    self,
    observation,
    info,
    action,
    reward,
    next_observation,
    next_info,
    terminated,
    truncated,
) -> None

Record one transition. Called once immediately after each step, in this argument order.

learn​

def learn(self) -> None

Perform one learning update (e.g. gradient step). Called once after each experience call.

save_models​

def save_models(self, save_to: Path) -> None

Persist agent state (weights, buffers, etc.) under the given directory. Training runs trigger saves periodically via RLExecutor.

load_models​

def load_models(self, load_from: Path) -> None

Restore agent state from the given directory at the start of each optimization run.

info

After save_models followed by load_models, external behaviour must match the saved state — the platform and your tests should see identical actions for identical observations.

Observation and action types depend on your environment; the ADK uses Any at the interface boundary. Gymnasium spaces on your env define the concrete shapes.