def build_agent_spaces(self) -> Tuple[Space, Space]:
"""Construct the action and observation spaces
Description of actions and observations:
https://github.com/google-research/football/blob/master/gfootball/doc/observation.md
""" # noqa: E501
action_space = Discrete(19)
# The football field's corners are [+-1., +-0.42]. However, the players
# and balls may get out of the field. Thus we multiply those limits by
# a factor of 2.
xlim = 1. * 2
ylim = 0.42 * 2
num_players: int = 11
xy_space = Box(
np.array([-xlim, -ylim], dtype=np.float32),
np.array([xlim, ylim], dtype=np.float32))
xyz_space = Box(
np.array([-xlim, -ylim, 0], dtype=np.float32),
np.array([xlim, ylim, np.inf], dtype=np.float32))
observation_space = DictSpace({
"controlled_players": Discrete(2),
"players_raw": TupleSpace([
DictSpace({
# ball information
"ball": xyz_space,
"ball_direction": Box(-np.inf, np.inf, (3, )),
"ball_rotation": Box(-np.inf, np.inf, (3, )),
"ball_owned_team": Discrete(3),
"ball_owned_player": Discrete(num_players + 1),
# left team
"left_team": TupleSpace([xy_space] * num_players),
"left_team_direction": TupleSpace(
[xy_space] * num_players),
"left_team_tired_factor": Box(0., 1., (num_players, )),
"left_team_yellow_card": MultiBinary(num_players),
"left_team_active": MultiBinary(num_players),
"left_team_roles": MultiDiscrete([10] * num_players),
# right team
"right_team": TupleSpace([xy_space] * num_players),
"right_team_direction": TupleSpace(
[xy_space] * num_players),
"right_team_tired_factor": Box(0., 1., (num_players, )),
"right_team_yellow_card": MultiBinary(num_players),
"right_team_active": MultiBinary(num_players),
"right_team_roles": MultiDiscrete([10] * num_players),
# controlled player information
"active": Discrete(num_players),
"designated": Discrete(num_players),
"sticky_actions": MultiBinary(10),
# match state
"score": Box(-np.inf, np.inf, (2, )),
"steps_left": Box(0, np.inf, (1, )),
"game_mode": Discrete(7)
})
])
})
return action_space, observation_space
def get_policy_configs_for_game(
game_name: str) -> Tuple[dict, Callable[[AgentID], PolicyID]]:
# The RLlib server must know about the Spaces that the Client will be
# using inside Unity3D, up-front.
obs_spaces = {
# 3DBall.
"3DBall":
Box(float("-inf"), float("inf"), (8, )),
# 3DBallHard.
"3DBallHard":
Box(float("-inf"), float("inf"), (45, )),
# GridFoodCollector
"GridFoodCollector":
Box(float("-inf"), float("inf"), (40, 40, 6)),
# Pyramids.
"Pyramids":
TupleSpace([
Box(float("-inf"), float("inf"), (56, )),
Box(float("-inf"), float("inf"), (56, )),
Box(float("-inf"), float("inf"), (56, )),
Box(float("-inf"), float("inf"), (4, )),
]),
# SoccerStrikersVsGoalie.
"Goalie":
Box(float("-inf"), float("inf"), (738, )),
"Striker":
TupleSpace([
Box(float("-inf"), float("inf"), (231, )),
Box(float("-inf"), float("inf"), (63, )),
]),
# Sorter.
"Sorter":
TupleSpace([
Box(float("-inf"), float("inf"), (
20,
23,
)),
Box(float("-inf"), float("inf"), (10, )),
Box(float("-inf"), float("inf"), (8, )),
]),
# Tennis.
"Tennis":
Box(float("-inf"), float("inf"), (27, )),
# VisualHallway.
"VisualHallway":
Box(float("-inf"), float("inf"), (84, 84, 3)),
# Walker.
"Walker":
Box(float("-inf"), float("inf"), (212, )),
# FoodCollector.
"FoodCollector":
TupleSpace([
Box(float("-inf"), float("inf"), (49, )),
Box(float("-inf"), float("inf"), (4, )),
]),
}
action_spaces = {
# 3DBall.
"3DBall": Box(float("-inf"), float("inf"), (2, ),
dtype=np.float32),
# 3DBallHard.
"3DBallHard": Box(float("-inf"),
float("inf"), (2, ),
dtype=np.float32),
# GridFoodCollector.
"GridFoodCollector": MultiDiscrete([3, 3, 3, 2]),
# Pyramids.
"Pyramids": MultiDiscrete([5]),
# SoccerStrikersVsGoalie.
"Goalie": MultiDiscrete([3, 3, 3]),
"Striker": MultiDiscrete([3, 3, 3]),
# Sorter.
"Sorter": MultiDiscrete([3, 3, 3]),
# Tennis.
"Tennis": Box(float("-inf"), float("inf"), (3, )),
# VisualHallway.
"VisualHallway": MultiDiscrete([5]),
# Walker.
"Walker": Box(float("-inf"), float("inf"), (39, )),
# FoodCollector.
"FoodCollector": MultiDiscrete([3, 3, 3, 2]),
}
# Policies (Unity: "behaviors") and agent-to-policy mapping fns.
if game_name == "SoccerStrikersVsGoalie":
policies = {
"Goalie":
(None, obs_spaces["Goalie"], action_spaces["Goalie"], {}),
"Striker":
(None, obs_spaces["Striker"], action_spaces["Striker"], {}),
}
def policy_mapping_fn(agent_id, episode, **kwargs):
return "Striker" if "Striker" in agent_id else "Goalie"
else:
policies = {
game_name:
(None, obs_spaces[game_name], action_spaces[game_name], {}),
}
def policy_mapping_fn(agent_id, episode, **kwargs):
return game_name
return policies, policy_mapping_fn