This repository provides a standalone version of Nash-DQN and Nash-DQN-with-Exploiter algorithm for two-player zero-sum Markov games, details see our paper: A Deep Reinforcement Learning Approach for Finding Non-Exploitable Strategies in Two-Player Atari Games. Zihan Ding, Dijia Su, Qinghua Liu, Chi Jin.
The Nash-DQN and Nash-DQN-with-Exploiter algorithms are also compared against other baselines methods like Self-play, Fictitious Self-play, Neural Fictitious Self-play, Policy Space Response Oracle, but in another library called MARS.
equilibrium_solver/ contains different solvers for Nash equilibrium, including ECOS, Nashpy, PuLP, CVXPY, Gurobipy, etc.solver_comparison.ipynb provides a time and solvability analysis for all solvers implemented in this repo;common/ contains all necessary RL components, including networks, env wrappers, buffers, training arguments, etc;agent.py is the base class of learning agents;dqn.py is the implementation of the DQN algorithm that Nash-DQN inherits;nash_dqn.py is the implementation of the Nash-DQN algorithm;nash_dqn_exploiter.py is the implementation of the Nash-DQN-with-Exploiter algorithm;launch.py is the launching script for train/test the algorithm;confs/default.yaml is the default configurations;confs/pettingzoo_boxing_v2_nash_dqn.yaml is an example configuration script for PettingZoo Boxing-v2 with Nash-DQN algorithm. Note that it will overwrite the same entries in the default configurations;confs/pettingzoo_boxing_v2_nash_dqn_exploiter.yaml is an example configuration script for PettingZoo Boxing-v2 with Nash-DQN-with-Exploiter algorithm. Note that it will overwrite the same entries in the default configurations;git clone https://github.com/quantumiracle/nash-dqn.git && cd nash-dqn
pip install -r requirements.txt
AutoROM --accept-licenseThe ECOS solver is needed, which can be installed via pip install ecos.
Others are general ML packages, like torch, tensorboard, numpy, gym, pettingzoo, supersuit, etc.
git clone https://github.com/quantumiracle/nash-dqn.git && cd nash-dqn
poetry install
poetry run AutoROM --accept-license
poetry run pip install torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu113
poetry run python launch.py --env pettingzoo_boxing_v2 --method nash_dqnTo train Nash-DQN/Nash-DQN-with-Exploiter on two-agent zero-sum game, boxing-v2 in Pettingzoo, run:
python launch.py --env pettingzoo_boxing_v2 --method nash_dqn
python launch.py --env pettingzoo_boxing_v2 --method nash_dqn_exploiter
Train with Wandb (for logging training results):
python launch.py --env pettingzoo_boxing_v2 --method nash_dqn --wandb_activate True --wandb_entity **YourWandbAccountName** --wandb_project **ProjectName**Test after training:
python launch.py --env pettingzoo_boxing_v2 --method nash_dqn --test True --load_model_idx './model/boxing_v2_NashDQN_2022-07-26-17-42-20/2000' --render TrueFull list of configurations:
Environment args:
* '--env', type=str, default=None, help='environment type and name'
* '--num_envs', type=int, default=1, help='number of environments for parallel sampling'
* '--ram', type=bool, default=False, help='use RAM observation'
* '--render', type=bool, default=False, help='render the scene'
* '--seed', type=str, default='random', help='random seed'
* '--record_video', type=bool, default=False, help='whether recording the video'
Agent args:
* '--algorithm', type=str, default=None, help='algorithm name'
* '--algorithm_spec.dueling', type=bool, default=False, help='DQN: dueling trick'
* '--algorithm_spec.replay_buffer_size', type=int, default=1e5, help='DQN: replay buffer size'
* '--algorithm_spec.gamma', type=float, default=0.99, help='DQN: discount factor'
* '--algorithm_spec.multi_step', type=int, default=1, help='DQN: multi-step return'
* '--algorithm_spec.target_update_interval', type=bool, default=False, help='DQN: steps skipped for target network update'
* '--algorithm_spec.eps_start', type=float, default=1, help='DQN: epsilon-greedy starting value'
* '--algorithm_spec.eps_final', type=float, default=0.001, help='DQN: epsilon-greedy ending value'
* '--algorithm_spec.eps_decay', type=float, default=5000000, help='DQN: epsilon-greedy decay interval'
Training args:
* '--num_process', type=int, default=1, help='use multiprocessing for both sampling and update'
* '--batch_size', type=int, default=128, help='batch size for update'
* '--max_episodes', type=int, default=50000, help='maximum episodes for rollout'
* '--max_steps_per_episode', type=int, default=300, help='maximum steps per episode'
* '--train_start_frame', type=int, default=0, help='start frame for training (not update when warmup)'
* '--method', dest='marl_method', type=str, default=None, help='method name'
* '--save_id', type=str, default='0', help='identification number for each run'
* '--optimizer', type=str, default='adam', help='choose optimizer'
* '--batch_size', type=int, default=128, help='batch size for update'
* '--learning_rate', type=float, default=1e-4, help='learning rate'
* '--device', type=str, default='gpu', help='computation device for model optimization'
* '--update_itr', type=int, default=1, help='number of updates per step'
* '--log_avg_window', type=int, default=20, help='window length for averaging the logged results'
* '--log_interval', type=int, default=20, help='interval for logging'
* '--test', type=bool, default=False, help='whether in test mode'
* '--exploit', type=bool, default=False, help='whether in exploitation mode'
* '--load_model_idx', type=str, default='0', help='index of the model to load'
* '--load_model_full_path', type=str, default='/', help='full path of the model to load'
* '--multiprocess', type=bool, default=False, help='whether using multiprocess or not'
* '--eval_models', type=bool, default=False, help='evalutation models during training (only for specific methods)'
* '--save_path', type=str, default='/', help='path to save models and logs'
* '--save_interval', type=int, default=2000, help='episode interval to save models'
* '--wandb_activate', type=bool, default=False, help='activate wandb for logging'
* '--wandb_entity', type=str, default='', help='wandb entity'
* '--wandb_project', type=str, default='', help='wandb project'
* '--wandb_group', type=str, default='', help='wandb project'
* '--wandb_name', type=str, default='', help='wandb name'
* '--net_architecture.hidden_dim_list', type=str, default='[128, 128, 128]', help='list of hidden dimensions for model'
* '--net_architecture.hidden_activation', type=str, default='ReLU', help='hidden activation function'
* '--net_architecture.output_activation', type=str, default=False, help='output activation function'
* '--net_architecture.hidden_activation', type=str, default='ReLU', help='hidden activation function'
* '--net_architecture.channel_list', type=str, default='[8, 8, 16]', help='list of channels for CNN'
* '--net_architecture.kernel_size_list', type=str, default='[4, 4, 4]', help='list of kernel sizes for CNN'
* '--net_architecture.stride_list', type=str, default='[2, 1, 1]', help='list of strides for CNN'
* '--marl_spec.min_update_interval', type=int, default=20, help='mininal opponent update interval in unit of episodes'
* '--marl_spec.score_avg_window', type=int, default=10, help='the length of window for averaging the score values'
* '--marl_spec.global_state', type=bool, default=False, help='whether using global observation' @article{ding2022deep,
title={A Deep Reinforcement Learning Approach for Finding Non-Exploitable Strategies in Two-Player Atari Games},
author={Ding, Zihan and Su, Dijia and Liu, Qinghua and Jin, Chi},
journal={arXiv preprint arXiv:2207.08894},
year={2022}
}