Added new hopper and swimmer examples

examples/hopper/MDP_funcs.py

@@ -12,14 +12,9 @@
 def make_train_MDP():
     return _standard_hopper()
 
-
 def make_test_MDP():
     return _heavy_hopper()
 
-
-def make_custom_MDP():
-    return _ensemble_hopper()
-
 # ===================================================================
 # Local functions to create envs
 
@@ -35,40 +30,17 @@ def _heavy_hopper():
     e.env.model.body_mass = bm; e.env.model.geom_size = gs;
     return e
 
-def _ensemble_hopper():
-    return GymEnv("UncertainHopper-v0")
-
-def _standard_walker():
-    return GymEnv('Walker2d-v1')
-
-def _heavy_walker():
-    # Make the torso heavy
-    e = GymEnv('Walker2d-v1')
-    bm = np.array(e.env.model.body_mass)
-    gs = np.array(e.env.model.geom_size)
-    bm[1] = 7; gs[1][0] = 0.1;
-    e.env.model.body_mass = bm; e.env.model.geom_size = gs;
-    return e
-
 # =======================================================================================
 # Generate environment corresponding to the given mode
 
 def get_environment(env_mode):
 
-    modes = ['train', 'test', 'custom', 'standard', 'heavy', 'random']    
+    modes = ['train', 'test']    
 
     if env_mode == 'train':
         env = make_train_MDP()
     elif env_mode == 'test':
         env = make_test_MDP()
-    elif env_mode == 'custom':
-        env = make_custom_MDP()
-    elif env_mode == 'standard':
-        env = _make_standard_MDP()
-    elif env_mode == 'heavy':
-        env = _make_heavy_MDP()
-    elif env_mode == 'random':
-        env = _make_random_MDP()
     else:
         print "ERROR: Unknown environment mode specified. Allowed modes are ", modes
 

examples/hopper/README.md

@@ -1,5 +1,5 @@
 ### Example 1
-In this example, we train a (32,32) net policy on the standard hopper environment and test it on a heavy hopper (twice the torso mass). This is just to show an example. Typically, we need to train longer (for approx 250 or 300 iterations).
+In this example, we train a (64,64) net policy on the standard hopper environment and test it on a heavy hopper (twice the torso mass). This is just to show an example. Typically, we need to train longer (for approx 250 or 300 iterations) and with a lower learning rate (max_kl=0.01).
 
 Here, we see that training on the standard hopper which has a torso mass of approx 3.5 (calculated using density=1000 in MuJoCo) and testing it on a hopper with heavier torso (mass=7, twice of standard) fails badly. However, if we train on an ensemble, the performance on test MDPs improve dramatically. See paper for more details.
 

examples/hopper/experiment_0/iterations/baseline_101.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S'\xcd\xe7\x98\xcb\xce|\xa9\xc0\xba\x13\xc8\x19,\xf0\x80\xc0\x9ej\x9f\xfd.\x84\\@e\x11"m\xa9\x99o@\x90y\x06^+\x07J@\xc0\x92v7\n\xa9q@\xd1\x84yo\xe3%D@:\xa1H\x83\xe2\x13=\xc0\x14\x13@\x86i\x0b@@*{\x8b\xa4\xf1\x001@\xf6\xea \xec\xdf\xbc\x00\xc0\xf9\xf6\xd4\x87Y\xdf\x92@\x1c6|IQ\xf7\x19\xc0\xf5\xf8\x9b\xc0\x15\x9ay\xc0\xf6\xddS\x08<J\x86\xc0\x8d\x05\xd87\xb6\x1cD\xc0\x99\xe1\xff``\rJ\xc0\xf92\x96FMI\x1a@/\x8e\xaa^,\xa8\x06\xc0\xf8F(\xda-\x9a\xf8?+\xaf\xd2\x8b\xde5\xed\xbf\x80^3\x9a\xa1W\xc2\xbf\x12\\U0dFY\xc0\xbc\x0eWV\x00\x118@\x00:\x83\x8d\xa6\x92\x00\xc0\xcaU\x135\x08!\xa5@'
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.

examples/hopper/experiment_0/iterations/baseline_11.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S'\x86\x1aGk\xb5\x18r\xc0\xa5b\x811\x92\x16Z\xc0\xf6T0vb?N@r\x86H=\xbc\x00N@\x0b\rpv\xdd\x16D@V*\xba\x92\xec89\xc0\x18\x8bo\xccZ\x04&@\x18v\xe2\x15\xa5\x7f\x16\xc0\x9a\x83h\xccw\xc5\x1c@\xaaX\xc9\xa4\'\x16\x0e@\x80IP3\xeeJ\xd2?2\xb4.\xde\x97\xe4^@}o\xb5\xc4\xb1\xe3\x90\xc0\xb2\xb2\x80\t\xf66H\xc0m\xf1"\xc6Jl_@6\x1aT\xc1y-%@\x1bu\x14\xe9{$ @d\xdf\xe4}\xc5\xee\xe9?|Y\x08\xbcX\xef\xf4\xbf \xb3\x0c\xd5H\xc9\xdc?\x002]\xfd\x13R\x87\xbf\x00\xe0\xe5\xb9\x14\xb0\xb2?\x17\xfa\x12%\x00\x14K@\x1e\xcf!\xad\xd5\x92K\xc0\x81\x11\xf4\xa9\xc1\xb0+@b\xe1\xfd\xf7\xab\xd9x@'
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.

examples/hopper/experiment_0/iterations/baseline_21.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S'D\xb8\xdf\xb0o\xf1\x89\xc0`\xcd\xce%myc\xc0\x91\x02\x91\x85\xf8\xf9Q@T\x01Z\xdd\xd1`I@\xab:\xf8\xb31\x92\x15@\x95 \xc06\xda\xa0W@\xab4\x8c]\r-/@#}${\xf5t$@%\x1c\xb7\x81L\xdf\x0f\xc0\xf2\x04\xc0cB~\x01\xc0p\xfa]*\x00i\xfc\xbfWN\n\x16| t@7\x92\x1f\x94z\xc0\x9f\xc0\xfa\xd6\xe8\xc7z\xb3a\xc0\xa6\x08\x13\xf2_X{\xc0\n\xf0\xdc\x9d\x88>A@\x85\xfb\xeeg\xbf\xc9?\xc0K\xc0\x80\xf5\x9cg\xfe\xbf.\xcf?k\x99\xa4\x02\xc0\xd8p\xb2\xcd\xa3D\xcd?\xa8\xaf\x87\xa1Q\xc9\xbf\xbf\x00\xd9\xbd\xb5\x00%\xcd?\xe8\xba(Y\x8a\xce\x07\xc0\xd8\xb7\xe3\xa0,\xe1\x05@\x00\x9e\xa6\x8c\xb1\xe3\xdc\xbf\x19\x8aV\xfe\x90\n\x8d@'
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.

examples/hopper/experiment_0/iterations/baseline_31.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S"+)K2f\xbf\xa5\xc0\xb7\xc2X1jit\xc0c\x8e\x8fU'\x9bI\xc0\x13\xa4t\xfd\xd5\xa2Y@?\xb0&\x87\x84b-@*TP3\x16\x03a@\xed\xf9<\x87\xac.0@\x01\x84\xfb\x9a\x19\x1c-\xc0\xc7V?`\xfbJ/@\xb4Z\x01U\xe9\x98\x1c@\xe8\xd2_\xd8\x99\xec\xd6\xbf<\xdd\x94K\xddV\x8f@(\x1e\xf0C\xfbL\x8c\xc0\xa7\x84\xe8\xb9(jx\xc0\x8f\xe7P\x0b\xfd\xb4i@\xb9`\xe0\x0f(\xca@@\xe8\xcd\x9e\x08\xc5\xa4A\xc0P\xb9#\x99%.\xd7\xbf\x8fC\xe9\xd6T\xdd\xe4?\xc0f#\x02\x95\xdf\xb9?(\x9cD~\xf1\xf7\xd3?\xc0*\xdfp\xc7\xc5\xd4?\x8fO\x98\xe9s\xf1S\xc0\xe2\xf19z\xdf\x195@\x00M\x0f\xfb\xaf\xf2\xfb\xbf7\xe2\x13\x85Q\x8b\xa2@"
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.

examples/hopper/experiment_0/iterations/baseline_41.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S"\xc2\xd2u\xe9v\xed\x9f\xc0$!\x08\xb9R\xa8\x80\xc0\x15W\xc1 \xa2\x8d9@p\x9b\x08\x8b\xec\xf4w@\x80\x0e\x84\xc7Et\t\xc0\xc3\x99\x83\xf1\x1c\xf1j@&\xde\xf4\xd1?q$@\xdb\xed\\2\x85i%\xc0\xbfv, \xd3:0@\xee\xf3-A\x1c%%@p\xc2\x8b\xbfwn\xfa\xbf\xea~@\xa7(,\x88@\xea\xf1\xb0\x03\xf4\xf8\x98\xc0'\xf4\x00G,@q\xc0m,MS\x84\x1db@,\xbe\xcdb,\x88\x0f@\x9d\x9f$+\xec\xc5J\xc0\x89\x90\x93\xd3f\xb3\n\xc0\x14\x88\xf8%\xfd\x98\xff\xbf`\x86\x9b\xbf\xba\x82\xee?$\xbf\xfeTn\xaf\xe1\xbf\x00\xacG[\x93\xa4\xb7?\xf61\x9e\x14\x9dy`\xc0\\Fi\xfe\xb4\x7f@@\x00\xcd\x97\xff2\xb9\x03\xc0F\x198yh\xeb\x9b@"
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.

examples/hopper/experiment_0/iterations/baseline_51.pickle

@@ -0,0 +1,53 @@
+ccopy_reg
+_reconstructor
+p0
+(crllab.baselines.linear_feature_baseline
+LinearFeatureBaseline
+p1
+c__builtin__
+object
+p2
+Ntp3
+Rp4
+(dp5
+S'_coeffs'
+p6
+cnumpy.core.multiarray
+_reconstruct
+p7
+(cnumpy
+ndarray
+p8
+(I0
+tp9
+S'b'
+p10
+tp11
+Rp12
+(I1
+(I26
+tp13
+cnumpy
+dtype
+p14
+(S'f8'
+p15
+I0
+I1
+tp16
+Rp17
+(I3
+S'<'
+p18
+NNNI-1
+I-1
+I0
+tp19
+bI00
+S'\x1b\xceP\xf2\xf4\xf2\xa5\xc0\xa6\xc1\x8b\xde\'\xd0\x83\xc0\xf4C\xef\xff\x1a$=@~\xc6-\xd0\xe3ri@\xdd\xc7U\x16\x07\x83-\xc0\x1e@\xfe\xca\x0c\x7fr@\x10B\x810\x02\xa0&@\xad\nek\xe2$A\xc0%V>O?\x9e@@\xc1<Lv\x96\xb74@(\xad\xd3\x89\xddk\xff\xbf\xec\x94`\xb0F\xc1\x8e@[av\xd0\xe5\xbd\x85\xc0\x92 \x0bW\xdd\x0b|\xc0\xa7MV\xb7\x8e\x1dk\xc09\xff\xcc\x18\x88\xf7+\xc0Z\x17\xa5\x9e\x8e)Q\xc0\x8b\x01\x97\xfa*\xd7"\xc0\xc6l\x060\x9b\xe6\xf6\xbfhs\x8e\x9e:=\xfd?\xe8g\xeb\xa1\xfaw\xd2\xbf\x80F\x15\xe8\'\x81\xd0?\nu\xdf\xc3\xf7\x8da\xc0x`\x84\xdcj(A@\x00\xadeMQ\x03\x05\xc0A\x19\x88\xe0._\xa3@'
+p20
+tp21
+bsS'_reg_coeff'
+p22
+F1e-05
+sb.