Learnable Agent for Video Game Playing

Naive DQN & Double DQN & Dueling DQN & A2C
This project mainly used to compare the performance of DQN-based algorithm and A2C algorithm in several games, include ['CartPole-v0', 'LunarLander-v2', 'Breakout-v0', 'CarRacing-v0', 'Pendulum-v0']

I will add the score curve diagrams and the instruction for how to run it in the later day (After I get a job 🤣 )

how to run

Requires Pytorch 1.7+ and Python 3.7 to run.
Install the requirements from requirements.txt before start.

train the game by DQN-based alg, -gidx denotes game index, -dqn denotes DQN type

cd DQN
python main.py -mode train -gidx 0 -dqn 0 -step 5000 -lr 1e-3

run the model

cd DQN
python main.py -mode run -gidx 0 -dqn 0

train the game by A2C alg, -gidx denotes game index, -update denotes network update frequency

cd A2C
python main.py -mode train -gidx 0 -step 5000 -lr 1e-3 -update 128

run the model

cd A2C
python main.py -mode run -gidx 0