17 Commits

Include renames

Author	SHA1	Message	Date
Mike Clark	deab7043fc	add eps to avoid NaN ... I tracked some NaN's I was getting down to here. It happens when std is a small number, then var is even smaller, and log_density=inf. There are some problems where the agent will learn to use small standard deviations because jittering movements have a high cost, this makes those more stable.	2017-11-04 16:24:54 +08:00
Mike Clark	212ff0bc70	moving epsilon outside softplus ... I think you intended this to be outside the softplus. The reason is that it should be applied just before the log to avoid `log(0)=inf`.e.g. - `log(softplus(-1000+1e-5))=log(0)=inf`. - `log(softplus(-1000)+1e-5)=log(1e-5)!=inf`. Also this fixes a NaN I had.	2017-11-01 10:52:23 +08:00
Shangtong Zhang	a59f2d2911	Benchmark DDPG with Roboschool	2017-10-27 23:09:24 -06:00
Shangtong Zhang	6682937b54	Support python3 and pytorch 0.2	2017-10-26 23:01:57 -06:00
Shangtong Zhang	1b6de16b4f	Fix a critical bug in DDPG	2017-10-15 10:05:33 -06:00
Shangtong Zhang	3ed6f2a9db	Log wall time	2017-10-14 16:38:51 -06:00
Shangtong Zhang	c5cbc10f94	Refactor DDPG	2017-10-06 11:16:21 -06:00
Shangtong Zhang	c773da7d08	Support shared stats for continuous A3C	2017-10-04 20:58:42 -06:00
Shangtong Zhang	3935dfc428	Unifying networks for continuous A3C and PPO	2017-10-04 11:03:02 -06:00
Shangtong Zhang	3adceb5284	Gaussian actor	2017-10-04 10:15:28 -06:00
Shangtong Zhang	27f54ed420	Single thread PPO	2017-10-04 09:33:35 -06:00
Shangtong Zhang	e60e9feecb	Refactor for HRA	2017-08-29 22:10:13 -06:00
Shangtong Zhang	d52182882a	Implementation of hybrid reward architecture	2017-08-29 21:52:26 -06:00
Shangtong Zhang	b161200f0f	Add BN layer	2017-08-02 14:35:51 -06:00
Shangtong Zhang	5116733f22	DDPG Pendulum	2017-08-01 10:52:14 -06:00
Shangtong Zhang	1be3b44999	Continuous A3C	2017-07-31 22:53:09 -06:00
Shangtong Zhang	ce504e2d0f	Major refactor	2017-07-26 18:18:49 -06:00