openai/multiagent-competition

Code for the paper "Emergent Complexity via Multi-agent Competition"

Status: Archive (code is provided as-is, no updates expected)

Competitive Multi-Agent Environments

This repository contains the environments for the paper Emergent Complexity via Multi-agent Competition

Dependencies

Use pip install -r requirements.txt to install dependencies. If you haven't used MuJoCo before, please refer to the installation guide. The code has been tested with the following dependencies:

• Python version 3.6
• OpenAI GYM version 0.9.1 with MuJoCo 1.31 support (use mujoco-py version 0.5.7)
• Tensorflow version 1.1.0
• Numpy version 1.12.1

Installing Package

After installing all dependencies, make sure gym works with support for MuJoCo environments. Next install gym-compete package as:

cd gym-compete
pip install -e .

Check install is successful by coming out of the directory and trying import gym_compete in python console. Some users might require a sudo pip install.

Trying the environments

Agent policies are provided for the various environments in folder agent-zoo. To see a demo of all the environments do:

bash demo_tasks.sh all

To instead try a single environment use:

bash demo_tasks.sh <task>

where <task> is one of: run-to-goal-humans, run-to-goal-ants, you-shall-not-pass, sumo-ants, sumo-humans and kick-and-defend

multiagent-competition's Issues

Sorry,I can't find the training code!Is there any code that I can train my own model?

when I run"bash demo_tasks.sh all",there are mistakes like"AttributeError: 'MultiAgentEnv' object has no attribute 'env_scene'"

Evolution - multiplexing and demultiplexing

http://randomprojectionai.blogspot.com/2017/12/data-and-information-muliplexing-and.html

Mujoco version 1.5

Please give support for Mujoco 1.5 version since the new MacBook do not support 1.3 version

No Module Named baselines

It looked like everything installed correctly. I ran the example script for starting the quadrant env and I get that there's no baseline. Any ideas?

$ bin/examine.py examples/hide_and_seek_quadrant.jsonnet examples/hide_and_seek_quadrant.npz

Inferred:
names: examples/hide_and_seek_quadrant.jsonnet examples/hide_and_seek_quadrant.npz
arguments: {}

Traceback (most recent call last):
File "bin/examine.py", line 91, in
main()
File "/Users/phillipwalker/opt/anaconda3/lib/python3.7/site-packages/click/core.py", line 764, in call
return self.main(*args, **kwargs)
File "/Users/phillipwalker/opt/anaconda3/lib/python3.7/site-packages/click/core.py", line 717, in main
rv = self.invoke(ctx)
File "/Users/phillipwalker/opt/anaconda3/lib/python3.7/site-packages/click/core.py", line 956, in invoke
return ctx.invoke(self.callback, **ctx.params)
File "/Users/phillipwalker/opt/anaconda3/lib/python3.7/site-packages/click/core.py", line 555, in invoke
return callback(*args, **kwargs)
File "bin/examine.py", line 54, in main
from ma_policy.load_policy import load_policy
File "/Users/phillipwalker/mujoco-worldgen/multi-agent-emergence-environments/ma_policy/load_policy.py", line 9, in
from ma_policy.ma_policy import MAPolicy
File "/Users/phillipwalker/mujoco-worldgen/multi-agent-emergence-environments/ma_policy/ma_policy.py", line 9, in
from baselines.common.distributions import make_pdtype
ModuleNotFoundError: No module named 'baselines'

Add licence

Since the repository is archived, would it be possible to add a licence (e.g. MIT or Apache-2) so people can reuse parts of the code? Thanks!

Syntax error in policy.py

I am facing a syntax error in line 75 of policy.py. The error prompt is pointing out that there is something wrong with using * as def __init__'s 3rd argument. Any help in this regard would be awesome. Thanks for reading through.

Gym 0.9.* is breaking this repo.

Changing requirements.txt to include gym==0.7.4 instead of just gym fixes issues that are breaking this codebase. gym is currently moving to MuJoCo 150 and introduced breaking changes specifically calling the new load_model_from_path added in MuJoCo 150.

On MacOs you might also need to run xcode-select --install if you're getting zlib errors.

How to draw perturb Force as shown in the demo?

After setting force (3x1) and torque (3x1) as well as point (3x1), I was able to calculate qfrc_applied through functions.mj_applyFT, which is of dimension (nvx1). Any hint on how to render this force via mujoco_py?

requirements.txt doesn't specifies the required version of the packages

Description

The version number of the various dependencies required for this project isn't specified. So, when one installs directly from requirements.txt using pip, then a non-compatible version of packages are installed.

mjpro and mujoco_py installation experience

I have tried to install this on Mac and linux system.
Here's my experience.
The existed mjpro150 and corresponding mujoco_py 1.50 can't support this repo. Mac doesn't support mjpro131.
So to run this, you need to install it on linux system.
Choose mjpro131, mujoco_py 0.5.7, gym 0.7.4, then I met the problem
Exception: glfw failed to initialize
Fatal Python error: Couldn't create autoTLSkey mapping.
Change main.py file, env.render() to 'env.render(600,600)', the environment can work.

Does the move_reward_weight is the annealing factor in the formula 1 in the paper?

Hi,

I want to implement the training process of the multiagent-competition, I am wondering whether move_reward_weight is the annealing factor in the formula 1 in the paper and how can I make it anneal to 0 in the specified timestamps during the training, thank you!

raise NotImplementedError

ｗｈｅｎ i run the code in
self.env_scene = MultiAgentScene(self._env_xml_path, self.n_agents)

This error occurred
raise NotImplementedError