Unity 机器学习代理工具包 (ML-Agents) 是一个开源项目,它使游戏和模拟能够作为训练智能代理的环境。
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

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