haoshengzou
|
498b55c051
|
ppo with batch also works! now ppo improves steadily, dqn not so stable.
|
2018-03-10 17:30:11 +08:00 |
|
haoshengzou
|
92894d3853
|
working on off-policy test. other parts of dqn_replay is runnable, but performance not tested.
|
2018-03-09 15:07:14 +08:00 |
|
haoshengzou
|
675057c6b9
|
interfaces for advantage_estimation. full_return finished and tested.
|
2018-02-27 14:11:52 +08:00 |
|
Dong Yan
|
f3aee448e0
|
add option to show the running result of cartpole
|
2018-02-24 10:53:39 +08:00 |
|
Dong Yan
|
2163d18728
|
fix the env -> self._env bug
|
2018-02-10 03:42:00 +08:00 |
|
haoshengzou
|
9f96cc2461
|
finish design and running of ppo and actor-critic. advantage estimation module is not complete yet.
|
2018-01-17 14:21:50 +08:00 |
|
haoshengzou
|
ed25bf7586
|
fixed the bugs on Jan 14, which gives inferior or even no improvement. mistook group_ndims. policy will soon need refactoring.
|
2018-01-17 11:55:51 +08:00 |
|
haoshengzou
|
983cd36074
|
finished all ppo examples. Training is remarkably slower than the version before Jan 13. More strangely, in the gym example there's almost no improvement... but this problem comes behind design. I'll first write actor-critic.
|
2018-01-15 00:03:06 +08:00 |
|
haoshengzou
|
fed3bf2a12
|
auto target network. ppo_cartpole.py run ok. but results is different from previous version even with the same random seed, still needs debugging.
|
2018-01-14 20:58:28 +08:00 |
|
haoshengzou
|
dfcea74fcf
|
fix memory growth and slowness caused by sess.run(tf.multinomial()), now ppo examples are working OK with slight memory growth (1M/min), which still needs research
|
2018-01-03 20:32:05 +08:00 |
|
haoshengzou
|
4333ee5d39
|
ppo_cartpole.py seems to be working with param: bs128, num_ep20, max_time500; manually merged Normal from branch policy_wrapper
|
2018-01-02 19:40:37 +08:00 |
|