A3C

Tensorflow implementation of A3C, both discrete & continuous action space.

The a3c.py provides 2 mode of a3c: discrete & continuous.

Discrete: 
    The space of actions is limited. I use CartPole-v0 for test.
Continuous: 
    The space of actions is unlimited, and the shape of action is usually a list. I use Pendulum-v0 for test.

You can change the mode in Config.py

mode = 'continuous'
# mode = 'discrete'
GAME = 'CartPole-v0' if mode == 'discrete' else 'Pendulum-v0'

Ways to get the action-dimension are different between discrete-mode & continuous-mode:

if mode == 'discrete':  # Note：The action_space of CartPole-v0 does not contain attribute 'shape'.
    N_A = env.action_space.n
elif mode == 'continuous':  # Note: The action of Pendulum-v0 is a list with shape (1,).
    N_A = env.action_space.shape[0]

The result on Pendulum-v0: