Issues of multi brains in self-play

Did current version of self-play supports multi brain training? In my project I have four roules: striker, midfielder,back and goalie in a soccer game, but I failed to train with self-play. When I try only one brain at the same time the error didn't show up any more.

Traceback:
Traceback (most recent call last):
File "/home/seonghyeon/.conda/envs/mlagents2/bin/mlagents-learn", line 8, in <module>
sys.exit(main())
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/learn.py", line 495, in main
run_cli(parse_command_line())
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/learn.py", line 491, in run_cli
run_training(run_seed, options)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/learn.py", line 329, in run_training
tc.start_learning(env_manager)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents_envs/timers.py", line 258, in wrapped
return func(*args, **kwargs)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/trainer_controller.py", line 207, in start_learning
n_steps = self.advance(env_manager)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents_envs/timers.py", line 258, in wrapped
return func(*args, **kwargs)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/trainer_controller.py", line 267, in advance
num_steps = env.advance()
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/env_manager.py", line 91, in advance
num_step_infos = self._process_step_infos(new_step_infos)
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/env_manager.py", line 108, in _process_step_infos
name_behavior_id, ActionInfo.empty()
File "/home/seonghyeon/.conda/envs/mlagents2/lib/python3.7/site-packages/mlagents/trainers/agent_processor.py", line 101, in add_experiences
action = stored_take_action_outputs["action"][idx]
IndexError: index 14 is out of bounds for axis 0 with size 14

 

works for me for two brains. check the code.

 

If I'm not wrong there is no need to edit scripts to enable self-play. The project could work well in one brain self play and non self play. So there should be no problem in codes. Maybe it's because I'm using continuous actions? According to traceback it's actions that encouter errors.

 

The problem is that agents sometimes get no raycasts observations cuz higher altitude than borders. So I think the error has nothing to do with self-play implementation. But I do found that self-play didn't work well with multi brains.

 

Hi, I found a way to avoid this problem.
Go to the root directory of the project(where you can see "ml-agents" folder)，open the command line and execute "pip3 install -e ./ml-agents-envs" and "pip3 install -e ./ml-agents".Train again,and the training program will not crash.