Multi-Agent Setup for one network about mindmaker HOT 2 OPEN

I understand what you mean by non-marl distributed training of a single network, unfortunately at this moment I don’t have a working example of this. I would start by scouring the OpenAI gym and Stablebaselines docs to see if its implemented in any examples there, which would likely mean the same can be done in Mindmaker using similar methods, you would probably need to modify the source a bit. This might be what your looking for
https://github.com/Rohan138/marl-baselines3
https://github.com/HumanCompatibleAI/adversarial-policies/blob/baa359420641b721aa8132d289c3318edc45ec74/src/aprl/envs/multi_agent.py

from mindmaker.

I am working on the same issue. Parallel training is the only missing feature for this plugin before we can do serious research with, compared to using Unity ML-agents or OpenAI Gym.

The problem seems that we need to get the action/observation tuple in UE first, then dispatch it to each agent there. A lot of implementations automatically dispatch the data to each environment, which means we only get a single element in UE, which isn't what we want.

There are a few ways I have tried so far, albeit unsuccessfully. If you have any ideas on how to implement it I'll be interested.

- Implementing a vectorized environment from StableBaselines3 library
Explanation here: https://stable-baselines3.readthedocs.io/en/master/guide/vec_envs.html

So I end up creating different environments from the DummyVecEnv following the guidelines:

# DummyVecEnv
    def make_env(rank, seed = 0):
        env = UnrealEnvWrap()
        env.seed(rank + seed)
        return env 

    env = DummyVecEnv([make_env(i) for i in range(n_agents)])

However I get the following error :

File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\threading.py", line 980, in _bootstrap_inner
self.run()
File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\threading.py", line 917, in run
self._target(*self._args, **self._kwargs)
File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\site-packages\socketio\server.py", line 731, in _handle_event_internal
r = server._trigger_event(data[0], namespace, sid, *data[1:])
File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\site-packages\socketio\server.py", line 756, in _trigger_event
return self.handlers[namespace]event
File "C:\Users\fioshirk\Desktop\MMS2\25.py", line 243, in receive
env = DummyVecEnv([make_env(i) for i in range(n_agents)])
File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\site-packages\stable_baselines3\common\vec_env\dummy_vec_env.py", line 26, in init
self.envs = [fn() for fn in env_fns]
File "C:\Users\fioshirk\Anaconda3\envs\Mindmaker\lib\site-packages\stable_baselines3\common\vec_env\dummy_vec_env.py", line 26, in
self.envs = [fn() for fn in env_fns

TypeError: 'UnrealEnvWrap' object is not callable

- Implementing a custom vectorized environment

Instead of relying on the SB vectorized environment, we can create our own. It would be an environment that can loop through a tuple of environments at each step.

class VectorizedEnvironment(gym.Env): 
    def __init__(self, make_env, n):
        self.envs = tuple(make_env() for _ in range(n))
        print("Environments in Vectorized Environment: ",self.envs)
        
[...]
  def step(self, action):
          global observations
          global reward
          global UEreward
          global UEdone
          global obsflag
          obsflag = 0
 
          # Loop through each environment:
          for env, a in zip(self.envs, action):
              print("Step for environment number ",a)
              # send actions to UE as they are chosen by the RL algorithm

However it seems impossible to pass a tuple for the observation space, the model requires a gym space instead (for example a box). So not sure how to tackle this one either.

I'll have a look at the examples Krumiaa sent and perhaps it would work better.

from mindmaker.