updating experiment B vecenv

best_configs/shail-etienne4_expA.json

@@ -0,0 +1,33 @@
+{
+    "experiment": "A",
+    "trainenv": {
+        "stop_on_collision": false,
+        "safe_actions_collision_method": "circle",
+        "abort_unsafe_collision_method": "circle"
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 500,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 3,
+        "activation": 0,
+        "option": 0
+    },
+    "value": {
+        "learning_rate": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 3,
+        "n_hidden_layers_global": 2,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 150,
+    "seed": 0
+}

shail-experiment.py

@@ -62,7 +62,7 @@ def training_function(config):
             ), options=option_list[config['policy']['option']],
             safe_actions_collision_method=config['trainenv']['safe_actions_collision_method'], 
             abort_unsafe_collision_method=config['trainenv']['abort_unsafe_collision_method'],
-            ) for _ in range(60)]
+            ) for _ in range(120)]
 
     elif config['experiment'] == 'B': 
         envs = sum([[SafeOptionsEnv(Setobs(
@@ -78,7 +78,7 @@ def training_function(config):
             ), options=option_list[config['policy']['option']],
             safe_actions_collision_method=config['trainenv']['safe_actions_collision_method'], 
             abort_unsafe_collision_method=config['trainenv']['abort_unsafe_collision_method'],
-            ) for _ in range(15)] for track in range(4)],[])
+            ) for _ in range(30)] for track in range(4)],[])
     
     else:
         raise NotImplementedError
@@ -142,7 +142,7 @@ def training_function(config):
         v_opt=v_opt,
         v_iters=config['value']['iterations_per_epoch'],
         epochs=config['train_epochs'],
-        rollout_episodes=60, 
+        rollout_episodes=120, 
         rollout_steps=60,
         gamma=0.99,
         gae_lambda=0.9,
@@ -185,11 +185,11 @@ if __name__ == '__main__':
                     'learning_rate': 3e-4, 
                     'learning_rate_decay': 1.0, 
                     'clip_ratio': 0.2, 
-                    'iterations_per_epoch': 100,
-                    'hidden_layer_size': tune.grid_search([20, 40]),
-                    'n_hidden_layers': tune.grid_search([2, 3]), 
+                    'iterations_per_epoch': tune.grid_search([250, 500, 750]),
+                    'hidden_layer_size': 40, #tune.grid_search([20, 40]),
+                    'n_hidden_layers': 3, #tune.grid_search([2, 3]), 
                     'activation':0,
-                    'option': tune.grid_search(list(range(len(option_list))))
+                    'option': 0, #tune.grid_search(list(range(len(option_list))))
                 },
                 'value': {
                     'learning_rate': 1e-3, 
@@ -199,8 +199,8 @@ if __name__ == '__main__':
                     'learning_rate': 1e-3, 
                     'weight_decay': 1e-4, 
                     'iterations_per_epoch': 100, 
-                    'n_hidden_layers_element': tune.grid_search([3,4]),
-                    'n_hidden_layers_global': tune.grid_search([1,2]),
+                    'n_hidden_layers_element': 3, #tune.grid_search([3,4]),
+                    'n_hidden_layers_global': 2, #tune.grid_search([1,2]),
                     'hidden_layer_size': 10, 
                     'activation': 0,
                 },