def vision_for(self, agent):
"""
Строит видение мира для каждого агента
:param agent: машинка, из которой мы смотрим
:return: список из модуля скорости машинки, направленного угла между направлением машинки
и направлением на центр и `agent.rays` до ближайших стен трека (запустите картинку, и станет совсем понятно)
"""
state = self.agent_states[agent]
vision = [
abs(state.velocity),
np.sin(angle(-state.position, state.heading))
]
extras = len(vision)
delta = pi / (agent.rays - 1)
start = rotate(state.heading, -pi / 2)
sectors = len(self.map)
for i in range(agent.rays):
# define ray direction
ray = rotate(start, i * delta)
# define ray's intersections with walls
vision.append(np.infty)
for j in range(sectors):
inner_wall = self.map[j - 1][0], self.map[j][0]
outer_wall = self.map[j - 1][1], self.map[j][1]
intersect = intersect_ray_with_segment((state.position, ray),
inner_wall)
intersect = abs(
intersect -
state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
intersect = intersect_ray_with_segment((state.position, ray),
outer_wall)
intersect = abs(
intersect -
state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
for obstacle in self.obs:
for j in range(len(obstacle)):
obstacle_wall = obstacle[j - 1], obstacle[j]
intersect = intersect_ray_with_segment(
(state.position, ray), obstacle_wall)
intersect = abs(
intersect -
state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
assert vision[-1] < np.infty, \
"Something went wrong: {}, {}".format(str(state), str(agent.chosen_actions_history[-1]))
assert len(vision) == agent.rays + extras, \
"Something went wrong: {}, {}".format(str(state), str(agent.chosen_actions_history[-1]))
return vision
def reward(self, state, collision):
"""
Вычисление награды агента, находящегося в состоянии state.
Эту функцию можно (и иногда нужно!) менять, чтобы обучить вашу сеть именно тем вещам, которые вы от неё хотите
:param state: текущее состояние агента
:param collision: произошло ли столкновение со стеной на прошлом шаге
:return reward: награду агента (возможно, отрицательную)
"""
a = np.sin(angle(-state.position, state.heading))
heading_score = self.HEADING_REWARD * np.tanh(-2 * a)
idle_penalty = 0 if abs(state.velocity) > self.MIN_SPEED else -self.IDLENESS_PENALTY
speeding_penalty = 0 if abs(state.velocity) < self.MAX_SPEED else -self.SPEEDING_PENALTY * abs(state.velocity)
collision_penalty = - max(abs(state.velocity), 0.1) * int(collision) * self.COLLISION_PENALTY
dist = np.infty
sectors = len(self.map)
for j in range(sectors):
inner_wall = self.map[j - 1][0], self.map[j][0]
outer_wall = self.map[j - 1][1], self.map[j][1]
intersect = intersect_ray_with_segment((state.position, state.heading), inner_wall)
intersect = abs(intersect - state.position) if intersect is not None else np.infty
if intersect < dist:
dist = intersect
intersect = intersect_ray_with_segment((state.position, state.heading), outer_wall)
intersect = abs(intersect - state.position) if intersect is not None else np.infty
if intersect < dist:
dist = intersect
dist_penalty = 0 if collision else - max(abs(state.velocity), 1) * self.COLLISION_PENALTY * np.exp(-dist)
return heading_score + collision_penalty \
+ idle_penalty + speeding_penalty + dist_penalty
def vision_for(self, agent):
"""
Строит видение мира для каждого агента
:param agent: машинка, из которой мы смотрим
:return: список из модуля скорости машинки, направленного угла между направлением машинки
и направлением на центр и `agent.rays` до ближайших стен трека (запустите картинку, и станет совсем понятно)
"""
state = self.agent_states[agent]
vision = [abs(state.velocity), np.sin(angle(-state.position, state.heading))]
extras = len(vision)
delta = pi / (agent.rays - 1)
start = rotate(state.heading, - pi / 2)
sectors = len(self.map)
for i in range(agent.rays):
# define ray direction
ray = rotate(start, i * delta)
# define ray's intersections with walls
vision.append(np.infty)
for j in range(sectors):
inner_wall = self.map[j - 1][0], self.map[j][0]
outer_wall = self.map[j - 1][1], self.map[j][1]
intersect = intersect_ray_with_segment((state.position, ray), inner_wall)
intersect = abs(intersect - state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
intersect = intersect_ray_with_segment((state.position, ray), outer_wall)
intersect = abs(intersect - state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
for obstacle in self.obs:
for j in range(len(obstacle)):
obstacle_wall = obstacle[j - 1], obstacle[j]
intersect = intersect_ray_with_segment((state.position, ray), obstacle_wall)
intersect = abs(intersect - state.position) if intersect is not None else np.infty
if intersect < vision[-1]:
vision[-1] = intersect
assert vision[-1] < np.infty, \
"Something went wrong: {}, {}".format(str(state), str(agent.chosen_actions_history[-1]))
assert len(vision) == agent.rays + extras, \
"Something went wrong: {}, {}".format(str(state), str(agent.chosen_actions_history[-1]))
return vision