xtma / simple-pytorch-rl Goto Github PK

Hi Xtma,

Again very nice code!

Here a small bug (dqn.py, line 89):
probs = agent.eval_net(state)

change to:
probs = self.eval_net(state)

Hi Xtima,

Maybe you should convert this:
action_list = [(i * 4 - 2,) for i in range(args.num_actions)]

into
action_list = [(i / args.num_actions * 4 - 2,) for i in range(args.num_actions)]

now your action_list contains values that are outside the range of the max values (ea. 4*args.num_actions = 16).

Hi Xtma!

Please note that you should do this:

action = action.clamp(-2.0, 2.0)

Instead of this:

action.clamp(-2.0, 2.0)

... if you want the actions to be within a range. However, if you use the range, training is not good:
without clamp:

with clamp: