Module moog.tasks.composite_task

Composite task.

Classes

class CompositeTask (*tasks, timeout_steps=inf)
Expand source code 
class CompositeTask(abstract_task.AbstractTask):
    """CompositeTask task.

    This combines multiple tasks at once, summing the rewards from each of them.
    This can be useful for example to have a predator/prey task where there are
    positive rewards for catching the prey and negative rewards for being caught
    by the predators.
    """

    def __init__(self, *tasks, timeout_steps=np.inf):
        """Constructor.

        Args:
            tasks: Tasks to compose. Reward will be the sum of the rewards from
                each of these tasks.
            timeout_steps: After this number of steps since reset, a reset is
                forced.
        """
        self._tasks = tasks
        self._timeout_steps = timeout_steps

    def reset(self, state, meta_state):
        for task in self._tasks:
            task.reset(state, meta_state)

    def reward(self, state, meta_state, step_count):
        """Compute reward."""
        reward = 0
        should_reset = step_count >= self._timeout_steps
        for task in self._tasks:
            task_reward, task_should_reset = task.reward(
                state, meta_state, step_count)
            reward += task_reward
            should_reset = should_reset or task_should_reset
        
        return reward, should_reset

CompositeTask task.

This combines multiple tasks at once, summing the rewards from each of them. This can be useful for example to have a predator/prey task where there are positive rewards for catching the prey and negative rewards for being caught by the predators.

Constructor.

Args

tasks
Tasks to compose. Reward will be the sum of the rewards from each of these tasks.
timeout_steps
After this number of steps since reset, a reset is forced.

Ancestors

Methods

def reward(self, state, meta_state, step_count)
Expand source code 
def reward(self, state, meta_state, step_count):
    """Compute reward."""
    reward = 0
    should_reset = step_count >= self._timeout_steps
    for task in self._tasks:
        task_reward, task_should_reset = task.reward(
            state, meta_state, step_count)
        reward += task_reward
        should_reset = should_reset or task_should_reset
    
    return reward, should_reset

Compute reward.

Inherited members