def patch_gym_spaces(self, u_env):
'''
For standardization, use gym spaces to represent observation and action spaces for Unity.
This method iterates through the multiple brains (multiagent) then constructs and returns lists of observation_spaces and action_spaces
'''
observation_spaces = []
action_spaces = []
for a in range(len(u_env.brain_names)):
brain = self._get_brain(u_env, a)
observation_shape = (brain.get_observable_dim()['state'],)
if brain.is_discrete():
dtype = np.int32
action_space = spaces.Discrete(brain.get_action_dim())
else:
dtype = np.float32
action_space = spaces.Box(low=0.0, high=1.0, shape=(1,), dtype=dtype)
observation_space = spaces.Box(low=0, high=1, shape=observation_shape, dtype=dtype)
set_gym_space_attr(observation_space)
set_gym_space_attr(action_space)
observation_spaces.append(observation_space)
action_spaces.append(action_space)
# set for singleton
u_env.observation_space = observation_spaces[0]
u_env.action_space = action_spaces[0]
return observation_spaces, action_spaces
def patch_gym_spaces(self, env):
r"""
For standardization, use gym spaces to represent observation and action spaces for Unity.
This method iterates through the multiple brains (multiagent) then constructs and returns lists of observation_spaces and action_spaces
:param env:
:return:
"""
observation_spaces = []
action_spaces = []
for brain_index in range(len(env.brain_names)):
brain = self._get_brain(env, brain_index)
# TODO: Logging
utils.describe(brain)
observation_shape = (brain.get_observable_dim()['state'],)
action_dim = (brain.get_action_dim(),)
if brain.is_discrete():
dtype = np.int32
action_space = spaces.Discrete(brain.get_action_dim())
else:
dtype = np.float32
action_space = spaces.Box(low=0.0, high=1.0, shape=action_dim, dtype=dtype)
observation_space = spaces.Box(low=0, high=1, shape=observation_shape, dtype=dtype)
utils.set_gym_space_attr(observation_space)
utils.set_gym_space_attr(action_space)
observation_spaces.append(observation_space)
action_spaces.append(action_space)
# set for singleton
env.observation_space = observation_spaces[0]
env.action_space = action_spaces[0]
return observation_spaces, action_spaces