Merge branch 'master' of github.com:vitchyr/rlkit

Author: vitchyr

docs/SkewFit.md

@@ -1,13 +1,20 @@
 # Skew-Fit
 Requires [multiworld](https://github.com/vitchyr/multiworld) to be installed:
 ```
-pip install git+https://github.com/vitchyr/multiworld.git@28ee206f60a45690d484737466b558abdef191ea
+pip install git+https://github.com/vitchyr/multiworld.git@f711cdb
 ```
 
 Implementation of Skew-Fit. For more information:
  - [Videos](https://sites.google.com/view/skew-fit)
  - [arXiv](https://arxiv.org/abs/1903.03698)
- 
+
+To reproduce the results, use these library versions as the performance seems to depend on the library version:
+ -  multiworld: f711cdb (git hash)
+ - python: 3.5.2
+ - torch: 0.4.1.post2
+ - mujoco_py: 1.50.1.59
+ - gym: 0.10.5
+
 Here are the results you should expect from each script.
 These plots are generated with [viskit](https://github.com/vitchyr/viskit) 
 with smoothing on.

examples/dqn_and_double_dqn.py

@@ -1,5 +1,5 @@
 """
-Run DQN on grid world.
+Run DQN on CartPole-v0.
 """
 
 import gym
@@ -19,8 +19,8 @@
 
 
 def experiment(variant):
-    expl_env = gym.make('CartPole-v0')
-    eval_env = gym.make('CartPole-v0')
+    expl_env = gym.make('CartPole-v0').env
+    eval_env = gym.make('CartPole-v0').env
     obs_dim = expl_env.observation_space.low.size
     action_dim = eval_env.action_space.n
 
@@ -71,12 +71,10 @@ def experiment(variant):
     algorithm.train()
 
 
-
-
 if __name__ == "__main__":
     # noinspection PyTypeChecker
     variant = dict(
-        algorithm="SAC",
+        algorithm="DQN",
         version="normal",
         layer_size=256,
         replay_buffer_size=int(1E6),
@@ -94,6 +92,6 @@ def experiment(variant):
             learning_rate=3E-4,
         ),
     )
-    setup_logger('name-of-experiment', variant=variant)
+    setup_logger('dqn-CartPole', variant=variant)
     # ptu.set_gpu_mode(True)  # optionally set the GPU (default=False)
     experiment(variant)