Author: Fu Qingxu, CASIA

Introduction

The Hybrid Multi-agent Playground (HMP) is an experimental framework designed for Reinforcement Learning (RL) researchers. Unlike any other framework which only isolates the TASKs from the framework, HMP also separates the ALGORITHMs from the framework to achieve excellent compatibility.

Any algorithm, from the most straightforward script-AI to sophisticated RL learner, is abstracted into a module inside ./ALGORITHM/*.

We also put effect to interface all kinds of multi-agent environments, including gym, SMAC, air combat, et.al. Other frameworks, such as pymarl2, can interface with HMP as well. The entire HMP can disguise as an RL environment in pymarl2. We make it happen by building a particular ALGORITHM module, which runs pymarl2 in a subprocess. This work is ongoing. Currently, HMP can link to a modified version of pymarl2.

Please star the root Github project. Your encouragement is extremely important to us as researchers: https://github.com/binary-husky/hmp2g

Archived code used in our AAAI papers: https://github.com/binary-husky/hmp2g/tree/aaai-conc.

Demo

Unreal-Engine-Based Simulation

Note: render client is not included in the docker image. For visualization, refer to https://github.com/binary-husky/uhmap-visual-tool git resp.

git pull && python main.py -c ZHECKPOINT/uhmap_hete10vs10/render_result.jsonc

Web Demo of DCA (AAAI Paper Version)

http://cloud.fuqingxu.top:11601/

Decentralized Collective Assult (Improved Version, more Difficult Than AAAI Paper Version)

git pull && python main.py -c ZHECKPOINT/50RL-55opp/test-50RL-55opp.jsonc
(Also see https://www.bilibili.com/video/BV1vF411M7N9/)

Anti-Invasion Interception

git pull && python main.py -c ZHECKPOINT/test-aii515/test-aii515.jsonc --skip 

Hazardous Cargo Transport

git pull && python main.py -c ZHECKPOINT/test-cargo50/test-cargo50.jsonc --skip

Decentralized Collective Assult (AAAI Paper Version)

git pull && python main.py -c ZHECKPOINT/test-50+50/test-50+50.jsonc --skip
git pull && python main.py -c ZHECKPOINT/test-100+100/test-100+100.jsonc --skip

Dependency

We use docker to solve dependency: SetupDocker.

Please do not run on WindowsOS (low efficiency), but if you have to, also refer to the last part of setup_docker for pip requirements list.

Introducing the Structure of HMP

HMP's General Framework Structure

HMP's Config System (How to experiment)

HMP aims to optimize the parameter control experience as a framework for researchers. One configuration file is all that is needed for the config insertion.

<1> How to Config:

We discard the command line method to control parameters; instead, the commented-JSON (JSONC) is used for experiment configuration. To run an experiment, just type:

python main.py --cfg Json_Experiment_Config_File.jsonc

<2> How to Add and Override A Parameter:

Parameters assigned and overridden in the JSON file are NOT passed via init functions layer by layer as other frameworks usually do; instead, at the start of the main.py, a special program defined in UTIL/config_args.py will directly INJECT the overridden parameters to the desired location.

We give an example to demonstrate how simple it is to add new parameters. Suppose we want to introduce HP into DCA, then an initial HP, let say HP_MAX need to be defined as a parameter. Then:

• Open MISSION/collective_assult/collective_assult_parallel_run.py. (You can create new file if you wish so.)
• (Step1, Define It !) In ScenarioConfig class add a new line writing HP_MAX=100. (You can create another class if you wish so.)
• (Step2, Use It !) Anywhere you want to use the HP_MAX, first from xxx.collective_assult_parallel_run import ScenarioConfig, then use the parameter by init_hp_of_some_agent = ScenarioConfig.HP_MAX.
• (Step3, Change It !) To override the default value HP_MAX=100 in JSON (e.g., in ./example_dca.jsonc), you just need to add a line in the field "MISSION.collective_assult_debug.collective_assult_parallel_run.py->ScenarioConfig", for example:

{
    ...... (other field)
    "MISSION.collective_assult_debug.collective_assult_parallel_run.py->ScenarioConfig": {
        "HP_MAX": 222,  # <------ add this!
        "random_jam_prob": 0.05,    # (other config override in ScenarioConfig)
        ......
    },
    ...... (other field)
}

• You need not to worry about the format. You can write {"HP_MAX": 222} or {"HP_MAX": "222"}. If the value is a bool, you can write {"Key1":true,"Key2":false} or {"Key1":"True", "Key2":"False"}. Both are OK.
• Be aware, in Step2, HP_MAX=100 defines HP_MAX as Int. If you want a float, please write HP_MAX=100.0. Overriding an Int with float will trigger an assert error.
• All Done! Say bye-bye to annoying args passing and kwargs passing!

<3> How to Deal with Parameter Dependency:

Our framework can fully support complicated parameter dependency. Some parameters are sometimes just Chained together. Changing one of them can lead to the change of another. E.g., Let the number of parallel envs (num_threads) be 32, and we test the performance of every test_interval episode. We wish to have relate them with test_interval = 8*num_threads, meaning that a test run is shot every 8 rounds of parallel env executions. This function can be satisfied by defining a Chained var structure:

num_threads = 32  # run N parallel envs,
# define test interval
test_interval = 8*num_threads
# define the Chains of test interval
test_interval_cv = ChainVar(lambda num_threads:8*num_threads, chained_with=['num_threads'])
# all done! You need to do nothing else!

After this, you can expect the following override (JSON config override) behaviors:

• Changing Neither in JSON, then both parameters use the default (num_threads = 32, test_interval = 8*32)
• Changing only num_threads in JSON, then test_interval is also forced to change according to test_interval=8*num_threads.
• Changing only test_interval in JSON, the Chain will not work, obey JSON override, nothing has higher priority than an explicit JSON override.
• Changing both JSON, the Chain will not work, both obey JSON override, and nothing has higher priority than an explicit JSON override.

For details, please refer to config.py and UTIL/config_args.py, it is straightforward to understand once you read an example of this.

<4> How to Recover Configuration's Auto Backup:

When the experiment starts, the Json config override will be stored in ZHECKPOINT/the-experiment-note-you-defined/experiment.json. If the experiment later produces surprising results, you can consistently reproduce it again using this config backup.

Task Runner

Task Runner (task_runner.py) only has three lines of important code:

# line 1
actions_list, self.info_runner = self.platform_controller.act(self.info_runner)
# line 2: 
obs, reward, done, info = self.envs.step(actions_list)
# line 3: 
self.info_runner = self.update_runner(done, obs, reward, info)

• self.platform_controller.act: Get action, block information access between teams (LINK to ARGORITHM), handle algorithm internal state loopback.

• self.envs.step: Multi-thread environment step (LINK to MISSION).
• self.update_runner: Prepare obs (for decision making) and reward (for driving RL algorithms) for the next step.

The Time Sequence of HMP

In general, the HMP task runner can operate two ways:

• (Deprecated due) self.align_episode = False: threads immediately restart at terminal state, threads do not wait for each other
• self.align_episode = True: threads pause at terminal state, waiting until all threads terminate, then reset. Please refer to Hmp Time Sequence.

MISSION

Please refer to MISSION README.

Execution Pool

We designed a parallel execution pool based on shared memory, the most efficient inter-process communication method possible. This parallel pool is only functional on Linux, therefore, when running on Windows, we will automatically switch to a backup pool using the pipe.

• The efficient execution pool is defined in UTIL/shm_pool.pyx
• The windows-compatible execution pool is defined in UTIL/win_pool.py

Both of them is initialized in main.py, and they share the same APIs:

smart_pool = SmartPool(...)

Furthermore, the pickle and the reverse-pickle process is slow for large NumPy arrays, we solve this problem by copying raw NumPy memory directly to the shared memory area without pickle and significantly improve the data transfer efficiency.

When dealing with a large number of parallel environments (100+), the process coordination of OS can slow down even shared-memory communication.

As a compromise, we design a folding mechanism to allow a single process to run N=fold parallel environments to relieve the burden of OS. When folding is disabled (default disabled), fold=1.

However, note that setting fold>1 will not accelerate the parallel FPS (and usually decrease the FPS) for a single experiment, but it allows you to run more experiments simultaneously on the server.

VHMAP, Visualization of HMP

VHMAP is a visualization component of HMP. VHMAP

It is unfortunate that all existing RL environments fail to provide a visual interface satisfying the following useful features:

• Allowing visualizing while training without slowing down the training server.
• Using as few resources as possible.
• Friendly to SSH users, faster than RDP and X server, which is notoriously slow.
• No dependency, even an Android with a browser can access it.
• Smooth, using the client's CPU and GPU to render instead of the server's.
• Simple. No verbose lines about lights, buffering, refresh, and bla.bla.bla about which we researchers never care.

VHMAP is just the answer, Features:

• Python interface simplified to the max
• Rendering on the client side, automatic frame insertion, and silky smooth frame rates
• Few server-side dependencies
• Very low server-side resource consumption
• Based on ThreeJs, drag and drop support, mobile touch screen support
• Support switching between perspective and projection views
• Playback support
• Use zlib to compress data streams, low network bandwidth requirement

Interface functions, operation introduction.

• Right mouse button to pan, left mouse button to rotate, scroll wheel to zoom
• Support touch screen. If your laptop or phone has a touch screen
• Rendering refresh rate is displayed in the upper left corner
• play fps: how many keyframes per second (less than the rendering refresh rate, then insert frames; greater than the rendering refresh rate, then the excess is invalid)
• pause: pause
• next frame: pause and switch the next frame
• previous frame: pause and switch the previous frame
• loop to start: play all data, go back to the first frame
• ppt step: play a frame at a very slow speed. Easy to record the screen, will be stuck for a few seconds after pressing
• use orthcam: switch the perspective view (object near large and far small)/projection view (similar to what you would use in engineering drawing)
• P.S. The first time you switch to the projection view, you need to use the mouse wheel to enlarge the screen

Quick Start

Dependency

We use docker to solve dependency: setup_docker. This project uses techniques such as shared memory for extreme training efficiency, as a cost, WindowsOS+GPU training is not well supported (using pipe IO for Windows compat).

For Windows (Not recommended, please do NOT run under Windows if possible), also refer to the last part of setup_docker for pip requirements list.

Please read setup_docker.md first, and then set up the container using:

$ docker run -itd   --name  hmp-$USER \
--net host \
--gpus all \
--shm-size=16G \
fuqingxu/hmp:latest

# Now inside the HMP container
$ su hmp # (switch the account Inside the HMP container, password: hmp)
$ cd ~   # (go to home directory)

AAAI 2022

1. All Default: Testing

git pull && python main.py -c ZHECKPOINT/test-50+50/test-50+50.jsonc --skip
git pull && python main.py -c ZHECKPOINT/test-100+100/test-100+100.jsonc --skip

When the testing starts, open the revealed URL for monitoring. The front end is done by JavaScript and ThreeJS.

--------------------------------
JS visualizer online: http://172.18.116.150:aRandomPort
JS visualizer online (localhost): http://localhost:aRandomPort
--------------------------------

2. All Default: Training

git pull && python main.py -c ZDOCS/examples/dca/example_dca.jsonc
git pull && python main.py -c ZDOCS/examples/dca/train_old_dca.jsonc

3. Change Settings

Launch with:

python main.py --cfg xx.json

IJCNN 2022

git pull && python main.py -c ZHECKPOINT/test-aii515/test-aii515.jsonc --skip 
git pull && python main.py -c ZHECKPOINT/test-cargo50/test-cargo50.jsonc --skip

Others

git pull && python main.py --cfg ZHECKPOINT/adca-demo/test.json
git pull && python main.py --cfg ZHECKPOINT/basic-ma-40-demo/test.json

Project Roadmap

If you are interested in something, you may continue to read:

    Handling parallel environment             -->   task_runner.py & shm_env.py

    The link between teams and diverse algorithms -->   multi_team.py

    Adding new env                            -->   MISSION.env_router.py

    Adding algorithm                          -->   ALGORITHM.example_foundation.py

    Configuring by writing py files           -->   config.py

    Configuring by JSONC                       -->   xx.jsonc

    colorful printing                         -->   colorful.py

    auto pip deployer                         -->   pip_find_missing.py

    efficient parallel execting               -->   shm_pool.pyx

    auto GPU selection                        -->   auto_gpu.py

    logging/plotting bridge            -->   mcom.py & mcom_rec.py

    experiment batch executor                 -->   mprofile.py

How to Add a New Environment (MISSION) in HMP

Please refer to MISSION README for more details.

• Make a new jsonc config file, using 'example.jsonc' as a template
• mkdir in MISSION, e.g. ./MISSION/bvr_sim, copy src code of the environment inside it.
• Open MISSION/env_router.py, add the path of environment's init function in env_init_function_ref, e.g.:

env_init_function_ref = {
    "bvr": ("MISSION.bvr_sim.init_env", "make_bvr_env"),
}   
# bvr is the final name that HMP recognizes, 
# MISSION.bvr_sim.init_env is a py file, 
# ScenarioConfig is a class

• Open MISSION/env_router.py, add the path of environment's configuration in import_path_ref

import_path_ref = {
    "bvr": ("MISSION.bvr_sim.init_env", 'ScenarioConfig'),
}   
# bvr will be the final name that HMP recognizes, 
# MISSION.bvr_sim.init_env is a py file, 
# make_bvr_env is a function

• Write your own ScenarioConfig. (refer to MISSION.bvr_sim.init_env.ScenarioConfig, as a template).
• Write your own env init function. (refer to MISSION.bvr_sim.init_env.make_bvr_env, as a template).

Papers Supported by HMP

@article{fu2022concentration,
    title        = {Concentration Network for Reinforcement Learning of Large-Scale Multi-Agent Systems},
    author       = {Fu, Qingxu and Qiu, Tenghai and Yi, Jianqiang and Pu, Zhiqiang and Wu, Shiguang},
    year         = 2022,
    month        = {Jun.},
    journal      = {Proceedings of the AAAI Conference on Artificial Intelligence},
    volume       = 36,
    number       = 9,
    pages        = {9341--9349},
    doi          = {10.1609/aaai.v36i9.21165},
    url          = {https://ojs.aaai.org/index.php/AAAI/article/view/21165}
}


@inproceedings{fu2022cooperation,
    title        = {A Cooperation Graph Approach for Multiagent Sparse Reward Reinforcement Learning},
    author       = {Fu, Qingxu and Qiu, Tenghai and Pu, Zhiqiang and Yi, Jianqiang and Yuan, Wanmai},
    year         = 2022,
    booktitle    = {2022 International Joint Conference on Neural Networks (IJCNN)},
    pages        = {1--8},
    doi          = {10.1109/IJCNN55064.2022.9891991}
}

@article{fu2022solving,
    title        = {Solving the Diffusion of Responsibility Problem in Multiagent Reinforcement Learning with a Policy Resonance Approach},
    author       = {Fu, Qingxu and Qiu, Tenghai and Yi, Jianqiang and Pu, Zhiqiang and Ai, Xiaolin and Yuan, Wanmai},
    year         = 2022,
    journal      = {arXiv preprint arXiv:2208.07753}
}