def __init__(
self,
agent_num,
game_name="pbeauty",
p=0.67,
reward_type="abs",
action_range=(-1.0, 1.0),
):
self.agent_num = agent_num
self.p = p
self.game_name = game_name
self.reward_type = reward_type
self.action_range = action_range
self.action_spaces = MASpace(
tuple(Box(low=-1.0, high=1.0, shape=(1,)) for _ in range(self.agent_num))
)
self.observation_spaces = MASpace(
tuple(Discrete(1) for _ in range(self.agent_num))
)
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.t = 0
self.rewards = np.zeros((self.agent_num,))
if not self.game_name in PBeautyGame.get_game_list():
raise EnvironmentNotFound(f"The game {self.game_name} doesn't exists")
if self.game_name == "pbeauty":
if (
not self.reward_type
in PBeautyGame.get_game_list()[self.game_name]["reward_type"]
):
raise RewardTypeNotFound(
f"The reward type {self.reward_type} doesn't exists"
)
def __init__(self):
self.env = gym.make("LunarLanderContinuous-v2")
self.agent_num = 2
self.observation_spaces = MASpace(
tuple(self.env.observation_space for _ in range(self.agent_num))
)
self.action_spaces = MASpace(
tuple([Box(low=-1.0, high=1.0, shape=(1,)) for _ in range(self.agent_num)])
)
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
def __init__(self, agent_num, height):
self.agent_num = agent_num
self.height = height
obs_dim = height * 4 + 1
self.action_range = [0, 1, 2]
self.action_spaces = MASpace(
tuple(
Box(low=0, high=2, shape=(1, ), dtype=np.int32)
for _ in range(self.agent_num)))
self.observation_spaces = MASpace(
tuple(
Box(low=0, high=1, shape=(obs_dim, ), dtype=np.int32)
for _ in range(self.agent_num)))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.grid = []
self.busy_n = {}
def __init__(self,
env,
agent_num=None,
action_space=None,
observation_space=None):
self.env = env
self.action_space = action_space
self.observation_space = observation_space
self.agent_num = agent_num
if self.agent_num is None:
if hasattr(self.env, "agent_num"):
self.agent_num = self.env.agent_num
if hasattr(self.env, "n"):
self.agent_num = self.env.n
if hasattr(self.env, "n_agents"):
self.agent_num = self.env.n_agents
# print('malib', self.agent_num, env.get_obs_size(), env, env.action_space, env.observation_sapce)
self.action_space = self.action_spaces = MASpace(
tuple(self.action_space for _ in range(self.agent_num)))
if self.observation_space is None:
obs_dim = self.env.get_obs_size()
self.observation_spaces = MASpace(
tuple(
gym.spaces.Box(low=-np.inf,
high=+np.inf,
shape=(obs_dim, ),
dtype=np.float32)
for _ in range(self.agent_num)))
else:
self.observation_spaces = MASpace(
tuple(self.observation_space for _ in range(self.agent_num)))
# if hasattr(self.env, "action_space") and hasattr(self.env, "observation_sapce"):
# self.observation_spaces = MASpace(tuple(gym.spaces.Box(low=-np.inf, high=+np.inf, shape=(obs_dim,), dtype=np.float32) for _ in range(self.agent_num)))
# elif hasattr(self.env, "action_spaces") and hasattr(self.env, "observation_spaces"):
# self.action_spaces = self.env.action_spaces
# self.observation_spaces = self.env.observation_spaces
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
def __init__(self):
self.agent_num = 1
self.x = 0
self.y = 50
self.theta = 0
self.v = 2
self.w = 0
self.stoch = 0
self.t = 0
obs_lows = np.array([0, 0, -1. * np.pi / 3., 2., -1.])
obs_highs = np.array([50, 100, np.pi / 3., 5., 1.])
self.observation_spaces = MASpace(
tuple([Box(low=obs_lows, high=obs_highs)]))
self.action_spaces = MASpace(
tuple([Box(low=-1., high=1., shape=(2, ))]))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.avs = [0] * 10
self.aws = [0] * 10
self.rewards = list(
) # a list that records rewards obtained in current episode, in sequential order.
def __init__(
self,
game_name,
agent_num,
action_num,
payoff=None,
repeated=False,
max_step=25,
memory=0,
discrete_action=True,
tuple_obs=True,
):
self.game_name = game_name
self.agent_num = agent_num
self.action_num = action_num
self.discrete_action = discrete_action
self.tuple_obs = tuple_obs
game_list = MatrixGame.get_game_list()
if not self.game_name in game_list:
raise EnvironmentNotFound(
f"The game {self.game_name} doesn't exists")
expt_num_agent = game_list[self.game_name]["agent_num"]
expt_num_action = game_list[self.game_name]["action_num"]
if expt_num_agent != self.agent_num:
raise WrongNumberOfAgent(f"The number of agent \
required for {self.game_name} is {expt_num_agent}")
if expt_num_action != self.action_num:
raise WrongNumberOfAction(f"The number of action \
required for {self.game_name} is {expt_num_action}")
self.action_spaces = MASpace(
tuple(
Box(low=-1.0, high=1.0, shape=(1, ))
for _ in range(self.agent_num)))
self.observation_spaces = MASpace(
tuple(Discrete(1) for _ in range(self.agent_num)))
if self.discrete_action:
self.action_spaces = MASpace(
tuple(Discrete(action_num) for _ in range(self.agent_num)))
if memory == 0:
self.observation_spaces = MASpace(
tuple(Discrete(1) for _ in range(self.agent_num)))
elif memory == 1:
self.observation_spaces = MASpace(
tuple(Discrete(5) for _ in range(self.agent_num)))
else:
self.action_range = [-1.0, 1.0]
self.action_spaces = MASpace(
tuple(
Box(low=-1.0, high=1.0, shape=(1, ))
for _ in range(self.agent_num)))
if memory == 0:
self.observation_spaces = MASpace(
tuple(Discrete(1) for _ in range(self.agent_num)))
elif memory == 1:
self.observation_spaces = MASpace(
tuple(
Box(low=-1.0, high=1.0, shape=(12, ))
for _ in range(self.agent_num)))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.t = 0
self.repeated = repeated
self.max_step = max_step
self.memory = memory
self.previous_action = 0
self.previous_actions = []
self.ep_rewards = np.zeros(2)
if payoff is not None:
payoff = np.array(payoff)
assert payoff.shape == tuple([agent_num] +
[action_num] * agent_num)
self.payoff = payoff
if payoff is None:
self.payoff = np.zeros(
tuple([agent_num] + [action_num] * agent_num))
if self.game_name == "coordination_0_0":
self.payoff[0] = [[1, -1], [-1, -1]]
self.payoff[1] = [[1, -1], [-1, -1]]
elif self.game_name == "coordination_same_action_with_preference":
self.payoff[0] = [[2, 0], [0, 1]]
self.payoff[1] = [[1, 0], [0, 2]]
elif self.game_name == "zero_sum_nash_0_1":
# payoff tabular of zero-sum game scenario. nash equilibrium: (Agenat1's action=0,Agent2's action=1)
self.payoff[0] = [[5, 2], [-1, 6]]
self.payoff[1] = [[-5, -2], [1, -6]]
elif self.game_name == "matching_pennies":
# payoff tabular of zero-sumgame scenario. matching pennies
self.payoff[0] = [[1, -1], [-1, 1]]
self.payoff[1] = [[-1, 1], [1, -1]]
elif self.game_name == "matching_pennies_3":
self.payoff[0] = [[[1, -1], [-1, 1]], [[1, -1], [-1, 1]]]
self.payoff[1] = [[[1, -1], [1, -1]], [[-1, 1], [-1, 1]]]
self.payoff[2] = [[[-1, -1], [1, 1]], [[1, 1], [-1, -1]]]
elif self.game_name == "prison_lola":
self.payoff[0] = [[-1, -3], [0, -2]]
self.payoff[1] = [[-1, 0], [-3, -2]]
elif self.game_name == "prison":
self.payoff[0] = [[3, 1], [4, 2]]
self.payoff[1] = [[3, 4], [1, 2]]
elif self.game_name == "stag_hunt":
self.payoff[0] = [[4, 1], [3, 2]]
self.payoff[1] = [[4, 3], [1, 2]]
elif self.game_name == "chicken": # snowdrift
self.payoff[0] = [[3, 2], [4, 1]]
self.payoff[1] = [[3, 4], [2, 1]]
elif self.game_name == "harmony":
self.payoff[0] = [[4, 3], [2, 1]]
self.payoff[1] = [[4, 2], [3, 1]]
elif self.game_name == "wolf_05_05":
self.payoff[0] = [[0, 3], [1, 2]]
self.payoff[1] = [[3, 2], [0, 1]]
# \alpha, \beta = 0, 0.9, nash is 0.5 0.5
# Q tables given, matian best response, learn a nash e.
elif self.game_name == "climbing":
self.payoff[0] = [[11, -30, 0], [-30, 7, 6], [0, 0, 5]]
self.payoff[1] = [[11, -30, 0], [-30, 7, 6], [0, 0, 5]]
elif self.game_name == "penalty":
self.payoff[0] = [[10, 0, 0], [0, 2, 0], [0, 0, 10]]
self.payoff[1] = [[10, 0, 0], [0, 2, 0], [0, 0, 10]]
elif self.game_name == "rock_paper_scissors":
self.payoff[0] = [[0, -1, 1], [1, 0, -1], [-1, 1, 0]]
self.payoff[1] = [[0, 1, -1], [-1, 0, 1], [1, -1, 0]]
self.rewards = np.zeros((self.agent_num, ))
def __init__(self,
game_name,
agent_num,
action_num,
state_num,
payoff=None,
transition=None):
self.game_name = game_name
self.agent_num = agent_num
self.action_num = action_num
self.state_num = state_num
game_list = StochasticMatrixGame.get_game_list()
if not self.game_name in game_list:
raise EnvironmentNotFound(
f"The game {self.game_name} doesn't exists")
expt_num_agent = game_list[self.game_name]['agent_num']
if expt_num_agent != self.agent_num:
raise WrongNumberOfAgent(f"The number of agent \
required for {self.game_name} is {expt_num_agent}")
expt_num_action = game_list[self.game_name]['action_num']
if expt_num_agent != self.action_num:
raise WrongNumberOfAction(f"The number of action \
required for {self.game_name} is {expt_num_action}")
expt_num_state = game_list[self.game_name]['state_num']
if expt_num_state != self.state_num:
raise WrongNumberOfState(f"The number of state \
required for {self.game_name} is {expt_num_state}")
self.action_spaces = MASpace(
tuple(
Box(low=-1., high=1., shape=(1, ))
for _ in range(self.agent_num)))
self.observation_spaces = MASpace(
tuple(Discrete(1) for _ in range(self.agent_num)))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.t = 0
if payoff is not None:
payoff = np.array(payoff)
assert payoff.shape == tuple([state_num, agent_num] +
[action_num] * agent_num)
self.payoff = payoff
if payoff is None:
self.payoff = np.zeros(
tuple([state_num, agent_num] + [action_num] * agent_num))
if transition is None:
self.transition = np.zeros(
tuple([state_num] + [action_num] * agent_num + [state_num]))
if self.game_name == 'PollutionTax':
self.payoff[0][0] = [[4., 3.], [7., 6.]]
self.payoff[0][1] = [[5., 8.], [4., 7.]]
self.payoff[1][0] = [[1., 0.], [4., 3.]]
self.payoff[1][1] = [[2., 5.], [1., 4.]]
self.transition[0] = [[[1., 0.], [0., 1.]], [[0., 1.], [0., 1.]]]
self.transition[1] = [[[1., 0.], [0., 1.]], [[0., 1.], [0., 1.]]]
elif self.game_name == 'three_matrix_games':
self.g1 = [[0., 3.], [2., -1.]]
self.g2 = [[0., 1.], [4., 3.]]
self.g = [['g1', 4.], [5., 'g2']]
self.rewards = np.zeros((self.agent_num, ))
self.state = 0
def __init__(
self,
world,
reset_callback=None,
reward_callback=None,
observation_callback=None,
info_callback=None,
done_callback=None,
shared_viewer=True,
):
self.world = world
self.agents = self.world.policy_agents
# set required vectorized gym env property
self.n = len(world.policy_agents)
# scenario callbacks
self.reset_callback = reset_callback
self.reward_callback = reward_callback
self.observation_callback = observation_callback
self.info_callback = info_callback
self.done_callback = done_callback
# environment parameters
self.discrete_action_space = True
# if true, action is a number 0...N, otherwise action is a one-hot N-dimensional vector
self.discrete_action_input = False
# if true, even the action is continuous, action will be performed discretely
self.force_discrete_action = (world.discrete_action if hasattr(
world, "discrete_action") else False)
# if true, every agent has the same reward
self.shared_reward = (world.collaborative if hasattr(
world, "collaborative") else False)
self.time = 0
# configure spaces
self.action_space = []
self.observation_space = []
obs_shapes = []
self.agent_num = len(self.agents)
for agent in self.agents:
total_action_space = []
# physical action space
if self.discrete_action_space:
u_action_space = spaces.Discrete(world.dim_p * 2 + 1)
else:
u_action_space = spaces.Box(
low=-agent.u_range,
high=+agent.u_range,
shape=(world.dim_p, ),
dtype=np.float32,
)
if agent.movable:
total_action_space.append(u_action_space)
# communication action space
if self.discrete_action_space:
c_action_space = spaces.Discrete(world.dim_c)
else:
c_action_space = spaces.Box(low=0.0,
high=1.0,
shape=(world.dim_c, ),
dtype=np.float32)
if not agent.silent:
total_action_space.append(c_action_space)
# total action space
if len(total_action_space) > 1:
# all action spaces are discrete, so simplify to MultiDiscrete action space
if all([
isinstance(act_space, spaces.Discrete)
for act_space in total_action_space
]):
act_space = MultiDiscrete(
[[0, act_space.n - 1]
for act_space in total_action_space])
else:
act_space = spaces.Tuple(total_action_space)
self.action_space.append(act_space)
else:
self.action_space.append(total_action_space[0])
# observation space
obs_dim = len(observation_callback(agent, self.world))
obs_shapes.append((obs_dim, ))
self.observation_space.append(
spaces.Box(low=-np.inf,
high=+np.inf,
shape=(obs_dim, ),
dtype=np.float32))
agent.action.c = np.zeros(self.world.dim_c)
# simpified for non-comm game
# self.action_spaces = MASpace(tuple(Box(low=-1., high=1., shape=(1,)) for _ in range(self.agent_num)))
# self.observation_spaces = MASpace(tuple(Discrete(1) for _ in range(self.agent_num)))
self.action_spaces = MASpace(
tuple(
Box(low=0.0, high=1.0, shape=(world.dim_p * 2 + 1, ))
for _ in range(self.agent_num)))
# print(obs_shapes)
self.observation_spaces = MASpace(
tuple(
Box(low=-np.inf, high=+np.inf, shape=obs_shape)
for obs_shape in obs_shapes))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.action_range = [0.0, 1.0]
# rendering
self.shared_viewer = shared_viewer
if self.shared_viewer:
self.viewers = [None]
else:
self.viewers = [None] * self.n
self._reset_render()
def __init__(self, game_name, agent_num, action_range=(-10, 10)):
self.game_name = game_name
self.agent_num = agent_num
self.action_range = action_range
game_list = DifferentialGame.get_game_list()
if not self.game_name in game_list:
raise EnvironmentNotFound(
f"The game {self.game_name} doesn't exists")
expt_num_agent = game_list[self.game_name]["agent_num"]
if expt_num_agent != self.agent_num:
raise WrongNumberOfAgent(f"The number of agent \
required for {self.game_name} is {expt_num_agent}")
self.action_spaces = MASpace(
tuple(
Box(low=-1.0, high=1.0, shape=(1, ))
for _ in range(self.agent_num)))
self.observation_spaces = MASpace(
tuple(
Box(low=-1.0, high=1.0, shape=(1, ))
for _ in range(self.agent_num)))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.t = 0
self.payoff = {}
if self.game_name == "zero_sum":
self.payoff[0] = lambda a1, a2: a1 * a2
self.payoff[1] = lambda a1, a2: -a1 * a2
elif self.game_name == "trigonometric":
self.payoff[0] = lambda a1, a2: np.cos(a2) * a1
self.payoff[1] = lambda a1, a2: np.sin(a1) * a2
elif self.game_name == "mataching_pennies":
self.payoff[0] = lambda a1, a2: (a1 - 0.5) * (a2 - 0.5)
self.payoff[1] = lambda a1, a2: (a1 - 0.5) * (a2 - 0.5)
elif self.game_name == "rotational":
self.payoff[0] = lambda a1, a2: 0.5 * a1 * a1 + 10 * a1 * a2
self.payoff[1] = lambda a1, a2: 0.5 * a2 * a2 - 10 * a1 * a2
elif self.game_name == "wolf":
def V(alpha, beta, payoff):
u = payoff[(0, 0)] - payoff[(0, 1)] - payoff[(1, 0)] + payoff[
(1, 1)]
return (alpha * beta * u + alpha *
(payoff[(0, 1)] - payoff[(1, 1)]) + beta *
(payoff[(1, 0)] - payoff[(1, 1)]) + payoff[(1, 1)])
payoff_0 = np.array([[0, 3], [1, 2]])
payoff_1 = np.array([[3, 2], [0, 1]])
self.payoff[0] = lambda a1, a2: V(a1, a2, payoff_0)
self.payoff[1] = lambda a1, a2: V(a1, a2, payoff_1)
elif self.game_name == "ma_softq":
h1 = 0.8
h2 = 1.0
s1 = 3.0
s2 = 1.0
x1 = -5.0
x2 = 5.0
y1 = -5.0
y2 = 5.0
c = 10.0
def max_f(a1, a2):
f1 = h1 * (-(np.square(a1 - x1) / s1) -
(np.square(a2 - y1) / s1))
f2 = h2 * (-(np.square(a1 - x2) / s2) -
(np.square(a2 - y2) / s2)) + c
return max(f1, f2)
self.payoff[0] = lambda a1, a2: max_f(a1, a2)
self.payoff[1] = lambda a1, a2: max_f(a1, a2)
else:
raise EnvironmentNotFound(
f"The game {self.game_name} doesn't exists")
self.rewards = np.zeros((self.agent_num, ))
def __init__(self,
n_pursuers,
n_evaders,
n_coop=2,
n_poison=10,
radius=0.015,
obstacle_radius=0.2,
obstacle_loc=np.array([0.5, 0.5]),
ev_speed=0.01,
poison_speed=0.01,
n_sensors=30,
sensor_range=0.2,
action_scale=0.01,
poison_reward=-1.,
food_reward=1.,
encounter_reward=.05,
control_penalty=-.5,
reward_mech='global',
addid=True,
speed_features=True,
**kwargs):
self.n_pursuers = n_pursuers
self.n_evaders = n_evaders
self.n_coop = n_coop
self.n_poison = n_poison
self.obstacle_radius = obstacle_radius
self.obstacle_loc = obstacle_loc
self.poison_speed = poison_speed
self.radius = radius
self.ev_speed = ev_speed
self.n_sensors = n_sensors
self.sensor_range = np.ones(self.n_pursuers) * sensor_range
self.action_scale = action_scale
self.poison_reward = poison_reward
self.food_reward = food_reward
self.control_penalty = control_penalty
self.encounter_reward = encounter_reward
self.n_obstacles = 1
self._reward_mech = reward_mech
self._addid = addid
self._speed_features = speed_features
self.seed()
self._pursuers = [
Archea(npu + 1,
self.radius * 3 / 4,
self.n_sensors,
self.sensor_range[npu],
addid=self._addid,
speed_features=self._speed_features)
for npu in range(self.n_pursuers)
]
self._evaders = [
Archea(nev + 1, self.radius * 3 / 4, self.n_pursuers,
self.sensor_range.mean() / 2)
for nev in range(self.n_evaders)
]
self._poisons = [
Archea(npo + 1, self.radius * 3 / 4, self.n_poison, 0)
for npo in range(self.n_poison)
]
self.observation_spaces = MASpace(
tuple(pursuer.observation_space for pursuer in self._pursuers))
self.action_spaces = MASpace(
tuple(pursuer.action_space for pursuer in self._pursuers))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
def __init__(self, world, reset_callback=None, reward_callback=None,
observation_callback=None, info_callback=None,
done_callback=None, shared_viewer=True):
self.world = world
self.world.discrete_action = True
self.agents = self.world.policy_agents
# set required vectorized gym env property
self.n = len(world.policy_agents)
# scenario callbacks
self.reset_callback = reset_callback
self.reward_callback = reward_callback
self.observation_callback = observation_callback
self.info_callback = info_callback
self.done_callback = done_callback
# environment parameters
self.discrete_action_space = True
# if true, action is a number 0...N, otherwise action is a one-hot N-dimensional vector
self.discrete_action_input = False
# if true, even the action is continuous, action will be performed discretely
self.force_discrete_action = world.discrete_action if hasattr(self.world, 'discrete_action') else False
# if true, every agent has the same reward
self.shared_reward = world.collaborative if hasattr(self.world, 'collaborative') else False
self.time = 0
# configure spaces
self.action_space = []
self.observation_space = []
obs_shapes = []
self.agent_num = len(self.agents)
## I think everything inside this loop is useless, self.action_spaces is used later and not self.action_space
for agent in self.agents:
total_action_space = []
# physical action space
if self.discrete_action_space:
u_action_space = spaces.Discrete((world.dim_p-1) * 2 + 3) ##
else:
u_action_space = spaces.Box(low=-agent.u_range, high=+agent.u_range, shape=(world.dim_p,), dtype=np.float32)
if agent.movable:
total_action_space.append(u_action_space)
# communication action space
if self.discrete_action_space:
c_action_space = spaces.Discrete(world.dim_c)
else:
c_action_space = spaces.Box(low=0.0, high=1.0, shape=(world.dim_c,), dtype=np.float32)
if not agent.silent:
total_action_space.append(c_action_space)
# total action space
if len(total_action_space) > 1:
# all action spaces are discrete, so simplify to MultiDiscrete action space
if all([isinstance(act_space, spaces.Discrete) for act_space in total_action_space]):
act_space = MultiDiscrete([[0, act_space.n - 1] for act_space in total_action_space])
else:
act_space = spaces.Tuple(total_action_space)
self.action_space.append(act_space)
else:
self.action_space.append(total_action_space[0])
# observation space
obs_dim = len(observation_callback(agent, self.world))
obs_shapes.append((obs_dim,))
self.observation_space.append(spaces.Box(low=-np.inf, high=+np.inf, shape=(obs_dim,), dtype=np.float32))
agent.action.c = np.zeros(self.world.dim_c)
# simpified for non-comm game
# self.action_spaces = MASpace(tuple(Box(low=-1., high=1., shape=(1,)) for _ in range(self.agent_num)))
# self.observation_spaces = MASpace(tuple(Discrete(1) for _ in range(self.agent_num)))
# action originally had 5 values - accel, +forcex, -forcex, +forcey, -forcey
# I added extra 2 components, change in rotation angle and shoot
self.action_spaces = MASpace(tuple(Box(low=0., high=1., shape=((world.dim_p-1) * 2 + 3,)) for _ in range(self.agent_num))) ##
self.observation_spaces = MASpace(tuple(Box(low=-np.inf, high=+np.inf, shape=obs_shape) for obs_shape in obs_shapes))
self.env_specs = MAEnvSpec(self.observation_spaces, self.action_spaces)
self.action_range = [0., 1.]
# rendering
self.shared_viewer = shared_viewer
if self.shared_viewer:
self.viewers = [None]
else:
self.viewers = [None] * self.n
mixer.init()
soundFiles = gym_fortattack.__file__[:-11]+'envs/Game/'
# bulletFile = os.path.realpath(__file__)[:-13]+'Game/bullet.mp3'
mixer.music.load(soundFiles+'bullet.mp3')
# print(gym_fortattack.__file__)
# time.sleep(5)
self.prevShot, self.shot = False, False # used for rendering
self._reset_render()
