AdditiveActionEnv

class AdditiveActionEnv(OptimizationEnv)

Additive-action environment: each step adds the action to current world controls without clipping (controls may exceed bounds; an exceed_world_control_bounds_reward penalty applies). Suited to multi-step episodes where actions are small deltas.

Registered as "AI4EE-Additive-Action-Env" (also the ADK DEFAULT_ENV). For clipping behavior, see ClippedAdditiveActionEnv ("AI4EE-Clipped-Additive-Action-Env").

Import

import adk  # registers Gymnasium env IDs
import gymnasium
from adk.executors.rl import EnvData

env = gymnasium.make("AI4EE-Additive-Action-Env", env_data=env_data)

Key differences from ClippedAdditiveActionEnv

	AdditiveActionEnv	ClippedAdditiveActionEnv
Out-of-bounds controls	Allowed; penalty via exceed_world_control_bounds_reward	Clipped to bounds
Default ID	AI4EE-Additive-Action-Env	AI4EE-Clipped-Additive-Action-Env

Shared constructor knobs include n_action_intervals, episode_maximum_steps, reward scaling, and observation/action space layout (see ClippedAdditiveActionEnv for the same reward and space semantics).

exceed_world_control_bounds_reward

exceed_world_control_bounds_reward: float = -50.0

Penalty when an additive action would push world controls outside bounds (unclipped env only).

target_generators

warning

Not implemented in either additive env variant — reset() raises NotImplementedError if set.

Methods

step

def step(self, action: NDArray[numpy.float32]) -> tuple[NDArray[numpy.float32], float, bool, bool, dict[str, Any]]

Same Gymnasium contract as ClippedAdditiveActionEnv.step: pass a flattened action array; receive observation, reward, terminated, truncated, and info.

reset / render

Same signatures and behavior as ClippedAdditiveActionEnv.