Change training schedule

Author: yura-hb

diploma_thesis/a.txt

@@ -18,8 +18,8 @@ graph: &graph
     is_machine_set_in_work_center_connected: False
     is_work_center_set_in_shop_floor_connected: False
 
-default_mods: &default_mods 
-  - ['util/infrastructure/cuda.yml']
+default_mods: &default_mods []
+#  - ['util/infrastructure/cuda.yml']
 
 ###############################################################################################
 

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/experiments/2 (DQN)/experiment.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'epsilon_greedy'
+                parameters:
+                  epsilon: 0.05
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -41,7 +43,7 @@ parameters:
                     kind: 'epsilon_greedy'
                     parameters:
                       epsilon: 0.4
-                      min_epsilon: 0.025
+                      min_epsilon: 0.05
                       decay_factor: 0.995
                       decay_steps: 30
 
@@ -51,7 +53,7 @@ parameters:
     kind: 'dqn'
     parameters:
       decay: 1.0
-      update_steps: 50
+      update_steps: 20
 
       memory:
         kind: 'replay'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/flexible_machine.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'epsilon_greedy'
+                parameters:
+                  epsilon: 0.05
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -41,7 +43,7 @@ parameters:
                     kind: 'epsilon_greedy'
                     parameters:
                       epsilon: 0.4
-                      min_epsilon: 0.025
+                      min_epsilon: 0.05
                       decay_factor: 0.995
                       decay_steps: 5
 
@@ -52,7 +54,7 @@ parameters:
     kind: 'dqn'
     parameters:
       decay: 1.0
-      update_steps: 50
+      update_steps: 20
 
       memory:
         kind: 'replay'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/flexible_marl_machine.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'sample'
+                parameters:
+                  is_distribution: False
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -49,7 +51,7 @@ parameters:
       device: 'cpu'
       sample_count: 96
       policy_step_ratio: 0.2
-      entropy_regularization: 0.01
+      entropy_regularization: 0.001
       rollback_ratio: 0.00
       critic_weight: 0.5
 
@@ -65,10 +67,9 @@ parameters:
 
       optimizer:
         model:
-          kind: 'adam'
+          kind: 'sgd'
           parameters:
-            lr: 0.0002
-            eps: 0.00001
+            lr: 0.001
 
       memory:
         kind: 'prioritized_replay'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/flexible_ppo_machine.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'epsilon_greedy'
+                parameters:
+                  epsilon: 0.05
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -51,7 +53,7 @@ parameters:
     kind: 'dqn'
     parameters:
       decay: 1.0
-      update_steps: 50
+      update_steps: 20
 
       memory:
         kind: 'replay'
@@ -65,7 +67,7 @@ parameters:
 
       optimizer:
         model:
-          kind: 'adam'
+          kind: 'sgd'
           parameters:
             lr: 0.001
 

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/machine.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'epsilon_greedy'
+                parameters:
+                  epsilon: 0.05
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -41,7 +43,7 @@ parameters:
                     kind: 'epsilon_greedy'
                     parameters:
                       epsilon: 0.4
-                      min_epsilon: 0.025
+                      min_epsilon: 0.05
                       decay_factor: 0.995
                       decay_steps: 5
 
@@ -52,7 +54,7 @@ parameters:
     kind: 'dqn'
     parameters:
       decay: 1.0
-      update_steps: 50
+      update_steps: 20
 
       memory:
         kind: 'replay'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/marl_machine.yml

@@ -19,7 +19,9 @@ parameters:
             kind: 'phase_selector'
             parameters:
               default:
-                kind: 'greedy'
+                kind: 'sample'
+                parameters:
+                  is_distribution: False
               phases:
                 - phase:
                     kind: 'warm_up'
@@ -61,14 +63,13 @@ parameters:
           reduction: 'none'
 
       value_loss:
-        kind: 'mse'
+        kind: 'huber'
 
       optimizer:
         model:
-          kind: 'adam'
+          kind: 'sgd'
           parameters:
-            lr: 0.0002
-            eps: 0.00001
+            lr: 0.001
 
       memory:
         kind: 'prioritized_replay'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/ppo_machine.yml

@@ -3,7 +3,7 @@
 kind: 'simulation'
 parameters:
   configuration:
-    timespan: 125000
+    timespan: 100000
     machines_per_work_center: 1
     work_center_count: 10
     deduce_naive_actions: True
@@ -23,7 +23,7 @@ parameters:
             uniform: [ 1, 50 ]
             noise: [ 0, 10 ]
         permutation:
-          uneveness: 5
+          uneveness: 3
         due_time:
           kind: 'uniform'
           parameters:

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/simulation.yml

@@ -5,14 +5,11 @@ layers:
   - kind: 'graph_model'
     parameters:
       layers:
-        - kind: 'gat'
+        - kind: 'gin'
           parameters:
             in_channels: -1
             out_channels: 128
             num_layers: 2
-            v2: True
-            heads: 2
-            concat: True
             hidden_channels: 128
 
       signature: 'graph -> actions_hidden, actions_batch'
@@ -37,17 +34,12 @@ layers:
         - kind: 'linear'
           parameters:
             dim: 128
-            activation: 'leaky_relu'
+            activation: 'relu'
             signature: 'x -> x'
         - kind: 'linear'
           parameters:
             dim: 128
-            activation: 'leaky_relu'
-            signature: 'x -> x'
-        - kind: 'linear'
-          parameters:
-            dim: 128
-            activation: 'leaky_relu'
+            activation: 'relu'
             signature: 'x -> x'
 
   - kind: 'linear'
@@ -61,14 +53,11 @@ layers:
   - kind: 'graph_model'
     parameters:
       layers:
-        - kind: 'gat'
+        - kind: 'gin'
           parameters:
             in_channels: -1
             out_channels: 128
             num_layers: 2
-            v2: True
-            heads: 2
-            concat: True
             hidden_channels: 128
 
       signature: 'graph -> values_hidden, values_batch'
@@ -85,17 +74,12 @@ layers:
   - kind: 'linear'
     parameters:
       dim: 128
-      activation: 'leaky_relu'
-      signature: 'values_hidden -> values_hidden'
-  - kind: 'linear'
-    parameters:
-      dim: 128
-      activation: 'leaky_relu'
+      activation: 'relu'
       signature: 'values_hidden -> values_hidden'
   - kind: 'linear'
     parameters:
       dim: 128
-      activation: 'leaky_relu'
+      activation: 'relu'
       signature: 'values_hidden -> values_hidden'
 
   - kind: 'output'

diploma_thesis/configuration/experiments/jsp/GRAPH-NN/templates/1/djsp/model.yml

@@ -5,14 +5,11 @@ layers:
   - kind: 'graph_model'
     parameters:
       layers:
-        - kind: 'gat'
+        - kind: 'gin'
           parameters:
             in_channels: -1
             out_channels: 128
             num_layers: 2
-            v2: True
-            heads: 2
-            concat: True
             hidden_channels: 128
 
       signature: 'graph -> actions_hidden, actions_batch'
@@ -37,17 +34,12 @@ layers:
         - kind: 'linear'
           parameters:
             dim: 128
-            activation: 'leaky_relu'
+            activation: 'relu'
             signature: 'x -> x'
         - kind: 'linear'
           parameters:
             dim: 128
-            activation: 'leaky_relu'
-            signature: 'x -> x'
-        - kind: 'linear'
-          parameters:
-            dim: 128
-            activation: 'leaky_relu'
+            activation: 'relu'
             signature: 'x -> x'
 
   - kind: 'linear'
@@ -61,14 +53,11 @@ layers:
   - kind: 'graph_model'
     parameters:
       layers:
-        - kind: 'gat'
+        - kind: 'gin'
           parameters:
             in_channels: -1
             out_channels: 128
             num_layers: 2
-            v2: True
-            heads: 2
-            concat: True
             hidden_channels: 128
 
       signature: 'graph -> values_hidden, values_batch'
@@ -85,17 +74,12 @@ layers:
   - kind: 'linear'
     parameters:
       dim: 128
-      activation: 'leaky_relu'
-      signature: 'values_hidden -> values_hidden'
-  - kind: 'linear'
-    parameters:
-      dim: 128
-      activation: 'leaky_relu'
+      activation: 'relu'
       signature: 'values_hidden -> values_hidden'
   - kind: 'linear'
     parameters:
       dim: 128
-      activation: 'leaky_relu'
+      activation: 'relu'
       signature: 'values_hidden -> values_hidden'
 
   - kind: 'output'