Source code for igp2.planning.reward

from copy import copy
from typing import Dict, List
import logging
import numpy as np

from igp2.core.trajectory import StateTrajectory
from igp2.core.goal import Goal
from igp2.core.cost import Cost

logger = logging.getLogger(__name__)


class Reward:
    """ Used to calculate rollout rewards in MCTS. """

    def __init__(self,
                 time_discount: float = 0.99,
                 factors: Dict[str, float] = None,
                 default_rewards: Dict[str, float] = None):
        """ Initialise a new Reward class.

        Args:
            time_discount: Discounting factor for each time step.
            factors: Reward component factors.
            default_rewards: Default rewards for specific events during rollout.
        """
        self._factors = {
            "time": 1.0,
            "jerk": -0.1,
            "angular_velocity": -0.1,
            "curvature": -0.1,
            "coll": 1.,
            "term": 1.,
            "dead": 1.,
        } if factors is None else factors

        self._default_rewards = {
            "coll": -1,
            "term": -1,
            "dead": -1,
        } if default_rewards is None else default_rewards

        self.COMPONENTS = list(set(self._factors).union(set(self._default_rewards)))

        self._time_discount = time_discount
        self._components = None
        self._reward = None
        self.reset()

    def __call__(self, *args, **kwargs):
        return self._calc_reward(*args, **kwargs)

    def _calc_reward(self,
                     collisions: List["Agent"] = None,
                     alive: bool = True,
                     ego_trajectory: StateTrajectory = None,
                     goal: Goal = None,
                     depth_reached: bool = False
                     ) -> float:
        if collisions:
            self._reward = self._factors.get("coll", 1.) * self._default_rewards["coll"]
            self._components["coll"] = self._reward
            logger.debug(f"Ego agent collided with agent(s): {collisions}")
        elif not alive:
            self._reward = self._factors.get("dead", 1.) * self._default_rewards["dead"]
            self._components["dead"] = self._reward
            logger.debug(f"Ego died during rollout!")
        elif ego_trajectory is not None and goal is not None:
            trajectory_rewards = self.trajectory_reward(ego_trajectory, goal)
            self._reward = sum([self._factors[comp] * rew for comp, rew in trajectory_rewards.items()])
            self._components.update(trajectory_rewards)
            logger.debug(f"Goal reached!")
        elif depth_reached:
            self._reward = self._factors.get("term", 1.) * self._default_rewards["term"]
            self._components["term"] = self._reward
            logger.debug("Reached final rollout depth!")

        return self._reward

    def trajectory_reward(self, trajectory: StateTrajectory, goal: Goal) -> Dict[str, float]:
        """ Calculate reward components for a given trajectory. """
        costs = Cost()
        costs.trajectory_cost(trajectory, goal)

        return {
            "time": self._time_discount ** trajectory.duration,
            "jerk": costs.cost_components["jerk"],
            "angular_velocity": costs.cost_components["angular_velocity"],
            "curvature": costs.cost_components["curvature"]
        }

    def reset(self):
        """ Reset rewards to initialisation values. """
        self._reward = None
        self._components = {rew: None for rew in self.COMPONENTS}

    @property
    def default_rewards(self) -> Dict[str, float]:
        """ Default rewards for rollout events e.g. collision, termination, and death. """
        return self._default_rewards

    @property
    def reward(self) -> float:
        """ The last calculate reward. """
        return self._reward

    @property
    def reward_components(self) -> Dict[str, float]:
        """ The current reward components """
        return self._components

    @property
    def cost_components(self) -> Dict[str, float]:
        cost_components = copy(self._components)
        if self._components["time"] is not None:
            tc = cost_components["time"]
            cost_components["time"] = np.log(tc) / np.log(self._time_discount)
        return cost_components

    @property
    def time_discount(self) -> float:
        """ Discounting factor for time-to-goal reward component. """
        return self._time_discount

    @property
    def factors(self) -> Dict[str, float]:
        """ Reward component factors. """
        return self._factors