Paper

• DSAC Implementation using ChainerRL.
• Our Soft Actor Critic is based on ChainerRL implementation.

• continuous action space
• discrete action space

• pip install -r requirements.txt

• Training Soft Q Imitation Learning (SQIL) and DSAC

  python train_sqil.py [options]

  • --load-demo [dirname] : replay buffer of demonstrations
  • --absorb : with absorbing state wrapper
  • --reward_func : use not constant rewards but generated rewards by a reward function.

  e.g.) DSAC with absorbing state wrapper in AntBulletEnv-v0 (random seed = 1)

  • python train_sqil.py --env AntBulletEnv-v0 --load-demo demos/4_episode/AntBulletEnv-v0 --absorb --reward-func --seed 1

python >= 3.7 and please see requirements.txt

If you'd like to use GPU, please pip install cupy-cudaOO

​ In relation to your version of cuda OO, please see the webpage of cupy.