使用了ppo算法实现了gobigger的单智能体训练模型，多智能体方法未实现。 安装DI版本为DI_engine-0.2.2
需要修改库 修改位置： ** 1： *****\envs<>\Lib\site-packages\DI_engine-0.2.2-py3.8.egg\ding\utils\default_helper.py ** split_data_generator() 修改内容：460行-464行，修改 elif isinstance(v, dict)判断为字典类型后的运算，增加函数：gobigger_data_deal(dict_data)
elif isinstance(data[k], dict):
gobigger_data, gobigger_batch, gobigger_player_num_per_team = gobigger_data_deal(data[k])
learn_batch = {k1: v1[indices[i:i + split_size]] for k1, v1 in gobigger_data.items()}
learn_batch['batch'] = split_size
learn_batch['player_num_per_team'] = gobigger_player_num_per_team
batch[k] = learn_batch
# batch[k] = {k1: v1[indices[i:i + split_size]] for k1, v1 in data[k].items()}
else:
batch[k] = data[k][indices[i:i + split_size]]
yield batch
def gobigger_data_deal(dict_data): # 拆分出 batch 与 player_num_per_team
assert isinstance(dict_data, dict), '修改有误，obs和next_obs不都为dict类型{}'.format(type(a))
control_data = copy.deepcopy(dict_data)
gobigger_batch = control_data['batch']
gobigger_player_num_per_team = control_data['player_num_per_team']
del control_data['batch']
del control_data['player_num_per_team']
return control_data, gobigger_batch, gobigger_player_num_per_team