Self play hyperparameter improvements (#4063)

config/ppo/SoccerTwos.yaml

     self_play:
       save_steps: 50000
       team_change: 200000
-      swap_steps: 50000
+      swap_steps: 2000
       window: 10
       play_against_latest_model_ratio: 0.5
       initial_elo: 1200.0

config/ppo/StrikersVsGoalie.yaml

     self_play:
       save_steps: 50000
       team_change: 200000
-      swap_steps: 25000
+      swap_steps: 1000
       window: 10
       play_against_latest_model_ratio: 0.5
       initial_elo: 1200.0
     self_play:
       save_steps: 50000
       team_change: 200000
-      swap_steps: 100000
+      swap_steps: 4000
       window: 10
       play_against_latest_model_ratio: 0.5
       initial_elo: 1200.0

config/ppo/Tennis.yaml

   Tennis:
     trainer_type: ppo
     hyperparameters:
-      batch_size: 1024
-      buffer_size: 10240
+      batch_size: 2048
+      buffer_size: 20480
       learning_rate: 0.0003
       beta: 0.005
       epsilon: 0.2
     self_play:
       save_steps: 50000
       team_change: 100000
-      swap_steps: 50000
+      swap_steps: 2000
       window: 10
       play_against_latest_model_ratio: 0.5
       initial_elo: 1200.0

docs/Training-ML-Agents.md

       window: 10
       play_against_latest_model_ratio: 0.5
       save_steps: 50000
-      swap_steps: 50000
+      swap_steps: 2000
       team_change: 100000
 ```
 

ml-agents/mlagents/trainers/settings.py

         # Assign team_change to about 4x save_steps
         return self.save_steps * 5
 
-    swap_steps: int = 10000
+    swap_steps: int = 2000
     window: int = 10
     play_against_latest_model_ratio: float = 0.5
     initial_elo: float = 1200.0