How to calculate agent_obs_shape and global_obs_shape in configuration file of smac env?

[Monarchlee]

I am confused by the calculation of agent_obs_shape and global_obs_shape. The docs just give me the value without any process about these two variables but I want to train agents on a self-made map. The official code is below.

        cuda=True,
        multi_agent=True,
        continuous=False,
        model=dict(
            # (int) agent_num: The number of the agent.
            # For SMAC 3s5z, agent_num=8; for 2c_vs_64zg, agent_num=2.
            agent_num=agent_num,
            # (int) obs_shape: The shapeension of observation of each agent.
            # For 3s5z, obs_shape=150; for 2c_vs_64zg, agent_num=404.
            # (int) global_obs_shape: The shapeension of global observation.
            # For 3s5z, obs_shape=216; for 2c_vs_64zg, agent_num=342.
            agent_obs_shape=150,
            #global_obs_shape=216,
            global_obs_shape=295,
            # (int) action_shape: The number of action which each agent can take.
            # action_shape= the number of common action (6) + the number of enemies.
            # For 3s5z, obs_shape=14 (6+8); for 2c_vs_64zg, agent_num=70 (6+64).
            action_shape=14,
            # (List[int]) The size of hidden layer
            # hidden_size_list=[64],
        )

Could anyone tell me how to calculate agent_obs_shape and global_obs_shape? Also, is it related to the type of entity or the kind of algorithm?

[yifan123]

1. It is related to the type of entity
2. agent_obs

   DI-engine/dizoo/smac/envs/smac_env.py

   Line 1041 in f1f0b55

   if self.flatten_observation:

3. global_obs shape

   DI-engine/dizoo/smac/envs/smac_env.py

   Line 1198 in f1f0b55

   def get_state(self, is_opponent=False):

4. You can print the shape in smac env and then set the right value in config and run again

[Monarchlee]

Thanks a lot. I will try to print the shape then.