Awesome

- If you want high sample efficiency, please use qmix_high_sample_efficiency.yaml
- which uses 4 processes for training, slower but higher sample efficiency.
- Performance is *not* comparable of models trained with different number of processes. 

PyMARL2

Open-source code for Rethinking the Implementation Tricks and Monotonicity Constraint in Cooperative Multi-Agent Reinforcement Learning.

This repository is fine-tuned for StarCraft Multi-agent Challenge (SMAC). For other multi-agent tasks, we also recommend an optimized implementation of QMIX: https://github.com/marlbenchmark/off-policy.

StarCraft 2 version: SC2.4.10. difficulty: 7.

2022.10.10 update: add qmix_high_sample_efficiency.yaml, which uses 4 processes for training, slower but higher sample efficiency.

2021.10.28 update: add Google Football Environments [vdn_gfootball.yaml] (use `simple115 features`).

2021.10.4 update: add QMIX with attention (qmix_att.yaml) as a baseline for Communication tasks.

Finetuned-QMIX

There are so many code-level tricks in the Multi-agent Reinforcement Learning (MARL), such as:

• Value function clipping (clip max Q values for QMIX)
• Value Normalization
• Reward scaling
• Orthogonal initialization and layer scaling
• Adam
• Neural networks hidden size
• learning rate annealing
• Reward Clipping
• Observation Normalization
• Gradient Clipping
• Large Batch Size
• N-step Returns(including GAE($\lambda$) and Q($\lambda$) ...)
• Rollout Process Number
• $\epsilon$-greedy annealing steps
• Death Agent Masking

Related Works

• Implementation Matters in Deep RL: A Case Study on PPO and TRPO
• What Matters In On-Policy Reinforcement Learning? A Large-Scale Empirical Study
• The Surprising Effectiveness of MAPPO in Cooperative, Multi-Agent Games

Using a few of tricks above (bold texts), we enabled QMIX (qmix.yaml) to solve almost all hard scenarios of SMAC (Fine-tuned hyperparameters for each scenarios).

Re-Evaluation

Afterwards, we re-evaluate numerous QMIX variants with normalized the tricks (a general set of hyperparameters), and find that QMIX achieves the SOTA.

Communication

We also tested our QMIX-with-attention (qmix_att.yaml, $\lambda=0.3$, attention_heads=4) on some maps (from NDQ) that require communication.

Google Football

We also tested VDN (vdn_gfootball.yaml) on some maps (from Google Football). Specially, we use simple115 features to train the model (The Google Football original paper use complex CNN features). We did not test QMIX because this environment does not provide global status information.

Usage

PyMARL is WhiRL's framework for deep multi-agent reinforcement learning and includes implementations of the following algorithms:

Value-based Methods:

• QMIX: QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning
• VDN: Value-Decomposition Networks For Cooperative Multi-Agent Learning
• IQL: Independent Q-Learning
• QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning
• Qatten: Qatten: A general framework for cooperative multiagent reinforcement learning
• QPLEX: Qplex: Duplex dueling multi-agent q-learning
• WQMIX: Weighted QMIX: Expanding Monotonic Value Function Factorisation

Actor Critic Methods:

• COMA: Counterfactual Multi-Agent Policy Gradients
• VMIX: Value-Decomposition Multi-Agent Actor-Critics
• LICA: Learning Implicit Credit Assignment for Cooperative Multi-Agent Reinforcement Learning
• DOP: Off-Policy Multi-Agent Decomposed Policy Gradients
• RIIT: Rethinking the Implementation Tricks and Monotonicity Constraint in Cooperative Multi-Agent Reinforcement Learning.

Installation instructions

Install Python packages

# require Anaconda 3 or Miniconda 3
conda create -n pymarl python=3.8 -y
conda activate pymarl

bash install_dependecies.sh

Set up StarCraft II (2.4.10) and SMAC:

bash install_sc2.sh

This will download SC2.4.10 into the 3rdparty folder and copy the maps necessary to run over.

Set up Google Football:

bash install_gfootball.sh

Command Line Tool

Run an experiment

# For SMAC
python3 src/main.py --config=qmix --env-config=sc2 with env_args.map_name=corridor

# For Difficulty-Enhanced Predator-Prey
python3 src/main.py --config=qmix_predator_prey --env-config=stag_hunt with env_args.map_name=stag_hunt

# For Communication tasks
python3 src/main.py --config=qmix_att --env-config=sc2 with env_args.map_name=1o_10b_vs_1r

# For Google Football (Insufficient testing)
# map_name: academy_counterattack_easy, academy_counterattack_hard, five_vs_five...
python3 src/main.py --config=vdn_gfootball --env-config=gfootball with env_args.map_name=academy_counterattack_hard env_args.num_agents=4

The config files act as defaults for an algorithm or environment.

They are all located in src/config. --config refers to the config files in src/config/algs --env-config refers to the config files in src/config/envs

Run n parallel experiments

# bash run.sh config_name env_config_name map_name_list (arg_list threads_num gpu_list experinments_num)
bash run.sh qmix sc2 6h_vs_8z epsilon_anneal_time=500000,td_lambda=0.3 2 0 5

xxx_list is separated by ,.

All results will be stored in the Results folder and named with map_name.

Kill all training processes

# all python and game processes of current user will quit.
bash clean.sh

Citation

@article{hu2021rethinking,
      title={Rethinking the Implementation Tricks and Monotonicity Constraint in Cooperative Multi-Agent Reinforcement Learning}, 
      author={Jian Hu and Siyang Jiang and Seth Austin Harding and Haibin Wu and Shih-wei Liao},
      year={2021},
      eprint={2102.03479},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}