您最多选择25个主题
主题必须以中文或者字母或数字开头，可以包含连字符 (-)，并且长度不得超过35个字符
37 行
1.4 KiB
37 行
1.4 KiB
import numpy as np
|
|
from mlagents.trainers.buffer import AgentBuffer
|
|
from mlagents_envs.base_env import BehaviorSpec
|
|
from mlagents.trainers.trajectory import ObsUtil
|
|
|
|
|
|
def create_agent_buffer(
|
|
behavior_spec: BehaviorSpec, number: int, reward: float = 0.0
|
|
) -> AgentBuffer:
|
|
buffer = AgentBuffer()
|
|
curr_obs = [
|
|
np.random.normal(size=sen_spec.shape).astype(np.float32)
|
|
for sen_spec in behavior_spec.sensor_specs
|
|
]
|
|
next_obs = [
|
|
np.random.normal(size=sen_spec.shape).astype(np.float32)
|
|
for sen_spec in behavior_spec.sensor_specs
|
|
]
|
|
action_buffer = behavior_spec.action_spec.random_action(1)
|
|
action = {}
|
|
if behavior_spec.action_spec.continuous_size > 0:
|
|
action["continuous_action"] = action_buffer.continuous
|
|
if behavior_spec.action_spec.discrete_size > 0:
|
|
action["discrete_action"] = action_buffer.discrete
|
|
|
|
for _ in range(number):
|
|
for i, obs in enumerate(curr_obs):
|
|
buffer[ObsUtil.get_name_at(i)].append(obs)
|
|
for i, obs in enumerate(next_obs):
|
|
buffer[ObsUtil.get_name_at_next(i)].append(obs)
|
|
buffer["actions"].append(action)
|
|
for _act_type, _act in action.items():
|
|
buffer[_act_type].append(_act[0, :])
|
|
buffer["reward"].append(np.ones(1, dtype=np.float32) * reward)
|
|
buffer["masks"].append(np.ones(1, dtype=np.float32))
|
|
buffer["done"] = np.zeros(number, dtype=np.float32)
|
|
return buffer
|