-

config/ppo/GridWorld.yaml

       normalize: false
       hidden_units: 128
       num_layers: 1
-      vis_encode_type: simple
+      vis_encode_type: fully_connected
     reward_signals:
       extrinsic:
         gamma: 0.9

config/sac/Walker.yaml

       reward_signal_steps_per_update: 30.0
     network_settings:
       normalize: true
-      hidden_units: 512
-      num_layers: 4
+      hidden_units: 256
+      num_layers: 3
       vis_encode_type: simple
       goal_conditioning_type: none
     reward_signals:
     keep_checkpoints: 5
-    max_steps: 150000000
+    max_steps: 15000000
     time_horizon: 1000
     summary_freq: 30000

ml-agents/mlagents/trainers/sac/optimizer_torch.py

     # gradient_penalty_weight = 10.0
     z_size = 128
     alpha = 0.0005
-    mutual_information = 10#0.5
+    mutual_information = 100#0.5
     EPSILON = 1e-7
     initial_beta = 0.0
 
-        print("VARIATIONAL : Settings : strength:", self.STRENGTH, " use_actions:", self._use_actions, " mutual_information : ", self.mutual_information)
+        sigma_start = 0.5
+        print("VARIATIONAL : Settings : strength:", self.STRENGTH, " use_actions:", self._use_actions, " mutual_information : ", self.mutual_information, "Sigma_Start : ", sigma_start)
         # state_encoder_settings = settings
         state_encoder_settings = NetworkSettings(normalize=True, num_layers=1)
         if state_encoder_settings.memory is not None:
             self._encoder = NetworkBody(new_spec, state_encoder_settings)
 
         self._z_sigma = torch.nn.Parameter(
-                torch.ones((self.z_size), dtype=torch.float), requires_grad=True
+                sigma_start * torch.ones((self.z_size), dtype=torch.float), requires_grad=True
             )
         # self._z_mu_layer = linear_layer(
         #     state_encoder_settings.hidden_units,