def _initialize(self, env):
assert isinstance(env.observation_space, (Box, Discrete)) and isinstance(env.action_space, (Box, Discrete)), 'action_space and observation_space must be one of available_type'
# process observation
ObsSpace = env.observation_space
if isinstance(ObsSpace, Box):
self.vector_dims = [ObsSpace.shape[0] if len(ObsSpace.shape) == 1 else 0]
# self.obs_high = ObsSpace.high
# self.obs_low = ObsSpace.low
else:
self.vector_dims = [int(ObsSpace.n)]
if len(ObsSpace.shape) == 3:
self.obs_type = 'visual'
self.visual_dims = [list(ObsSpace.shape)]
else:
self.obs_type = 'vector'
self.visual_dims = []
self.vector_info_type = NamedTupleStaticClass.generate_obs_namedtuple(n_agents=self.n,
item_nums=1 if self.obs_type == 'vector' else 0,
name='vector')
self.visual_info_type = NamedTupleStaticClass.generate_obs_namedtuple(n_agents=self.n,
item_nums=1 if self.obs_type == 'visual' else 0,
name='vector')
# process action
ActSpace = env.action_space
if isinstance(ActSpace, Box):
assert len(ActSpace.shape) == 1, 'if action space is continuous, the shape length of action must equal to 1'
self.action_type = 'continuous'
self._is_continuous = True
self.a_dim = ActSpace.shape[0]
elif isinstance(ActSpace, Tuple):
assert all([isinstance(i, Discrete) for i in ActSpace]) == True, 'if action space is Tuple, each item in it must have type Discrete'
self.action_type = 'Tuple(Discrete)'
self._is_continuous = False
self.a_dim = int(np.asarray([i.n for i in ActSpace]).prod())
discrete_action_dim_list = [i.n for i in ActSpace]
else:
self.action_type = 'discrete'
self._is_continuous = False
self.a_dim = env.action_space.n
discrete_action_dim_list = [env.action_space.n]
if not self._is_continuous:
self.discrete_action_list = get_discrete_action_list(discrete_action_dim_list)
self.reward_threshold = env.env.spec.reward_threshold # reward threshold refer to solved
self.EnvSpec = SingleAgentEnvArgs(
obs_spec=ObsSpec(vector_dims=self.vector_dims,
visual_dims=self.visual_dims),
a_dim=self.a_dim,
is_continuous=self._is_continuous,
n_agents=self.n
)
def initialize_environment(self):
'''
初始化环境，获取必要的信息，如状态、动作维度等等
'''
self.behavior_names = list(self.env.behavior_specs.keys())
self.is_multi_agents = len(self.behavior_names) > 1
self.first_bn = self.behavior_names[0]
self.first_fbn = self.first_bn.replace('?', '_')
self.behavior_agents = defaultdict(int)
self.behavior_ids = defaultdict(dict)
self.vector_idxs = defaultdict(list)
self.vector_dims = defaultdict(list)
self.visual_idxs = defaultdict(list)
self.visual_dims = defaultdict(list)
self.a_dim = defaultdict(int)
self.discrete_action_lists = {}
self.is_continuous = {}
self.empty_actiontuples = {}
self.vector_info_type = {}
self.visual_info_type = {}
self.env.reset()
for bn, spec in self.env.behavior_specs.items():
d, t = self.env.get_steps(bn)
self.behavior_agents[bn] = len(d)
self.behavior_ids[bn] = d.agent_id_to_index
for i, shape in enumerate(spec.observation_shapes):
if len(shape) == 1:
self.vector_idxs[bn].append(i)
self.vector_dims[bn].append(shape[0])
elif len(shape) == 3:
self.visual_idxs[bn].append(i)
self.visual_dims[bn].append(list(shape))
else:
raise ValueError(
"shape of observation cannot be understood.")
self.vector_info_type[
bn] = NamedTupleStaticClass.generate_obs_namedtuple(
n_agents=self.behavior_agents[bn],
item_nums=len(self.vector_idxs[bn]),
name='vector')
self.visual_info_type[
bn] = NamedTupleStaticClass.generate_obs_namedtuple(
n_agents=self.behavior_agents[bn],
item_nums=len(self.visual_idxs[bn]),
name='visual')
action_spec = spec.action_spec
if action_spec.is_continuous():
self.a_dim[bn] = action_spec.continuous_size
self.discrete_action_lists[bn] = None
self.is_continuous[bn] = True
elif action_spec.is_discrete():
self.a_dim[bn] = int(
np.asarray(action_spec.discrete_branches).prod())
self.discrete_action_lists[bn] = get_discrete_action_list(
action_spec.discrete_branches)
self.is_continuous[bn] = False
else:
raise NotImplementedError(
"doesn't support continuous and discrete actions simultaneously for now."
)
self.empty_actiontuples[bn] = action_spec.empty_action(
n_agents=self.behavior_agents[bn])
if self.is_multi_agents:
self.behavior_controls = defaultdict(int)
for bn in self.behavior_names:
self.behavior_controls[bn] = int(bn.split('#')[0])
self.env_copys = self.behavior_agents[
self.first_bn] // self.behavior_controls[self.first_bn]