Skip to main content

RL Run Data

When you use RLExecutor, it builds two data objects from the optimization context before constructing your RLAgentEnv subclass:

  • EnvData — spaces, defaults, and simulator callbacks for environments
  • AgentData — optimization state the agent reads directly

These types live under adk.executors.rl and are not part of the core ADK contract. The core ADK parses the platform payload into OptimizationContext via BaseExecutor.build_optimization_context(); only RLExecutor maps that context into EnvData and AgentData. If you subclass BaseExecutor yourself, work with OptimizationContext instead. See Specifications and What To Do Next.

from adk.executors.rl import EnvData, AgentData, RLAgentEnv

class MyAgent(RLAgentEnv):
    def __init__(self, env_data: EnvData, agent_data: AgentData) -> None:
        super().__init__(env_data, agent_data)
        if agent_data.optimization_data.inference:
            ...

EnvData​

EnvData holds what most Gymnasium environment implementations need to turn an optimization specification into an RL problem:

  • World control spaces — static, optimized, and randomized controllable parameters (static_world_controls_space, optimized_world_controls_space, randomized_world_controls_space).

  • Default world controls — starting values for each parameter group.

  • World observation space — evaluated observations the environment exposes to the agent.

  • Targets space — target value ranges used during optimization.

  • Internal structure graph — optional graph of the system under optimization (when instrumented).

  • Default world features — optional internal features extracted from the system.

  • step_world — callback to apply design parameters and read observations from the simulator.

  • optimization_data — inference flag and loaded Genie model for the run.

RL agents typically pass env_data into gymnasium.make for a built-in environment or an OptimizationEnv subclass. See the EnvData API page for the full field list.

AgentData​

AgentData is the slimmer companion passed alongside EnvData. It carries information your agent needs that is not wired through the environment:

  • optimization_data — whether the run is in inference mode (using a learnt model) or training mode, and the loaded Genie model (hyper-parameters, metadata, target specifications, world control specifications).

The Genie model inside optimization_data reflects the model files under <models directory>/<name>/ selected for this optimization on the platform. See AgentData for the type definition.