add_policy and create_policy separated

UnitySDK/Assets/ML-Agents/Scripts/Policy/BehaviorParameters.cs

         //string m_BehaviorIdentifier;
 
         
-        private string concatBehaviorIdentifiers()
-        {
-           return m_BehaviorName + "?team=" + m_TeamID;
-        }
-
         public BrainParameters brainParameters
         {
             get { return m_BrainParameters; }
         {
             
-            get { return concatBehaviorIdentifiers();}
+            get { return m_BehaviorName + "?team=" + m_TeamID;} 
 
         }
 
                 case BehaviorType.Default:
                     if (FindObjectOfType<Academy>().IsCommunicatorOn)
                     {
-                        return new RemotePolicy(m_BrainParameters, concatBehaviorIdentifiers());
+                        return new RemotePolicy(m_BrainParameters, behaviorName);
                     }
                     if (m_Model != null)
                     {

ml-agents/mlagents/trainers/ppo/trainer.py

 import numpy as np
 
 from mlagents.envs.brain import BrainParameters, BrainInfo
+from mlagents.trainers.tf_policy import TFPolicy
 from mlagents.trainers.ppo.policy import PPOPolicy
 from mlagents.trainers.ppo.multi_gpu_policy import MultiGpuPPOPolicy, get_devices
 from mlagents.trainers.rl_trainer import RLTrainer, AllRewardsOutput
         self.policy = None
 
     def process_experiences(
-        self, current_info: BrainInfo, next_info: BrainInfo
+        self, name_behavior_id: str, current_info: BrainInfo, next_info: BrainInfo
+        :param name_behavior_id: string policy identifier.
         :param current_info: current BrainInfo.
         :param next_info: next BrainInfo.
         """
         self.clear_update_buffer()
         self.trainer_metrics.end_policy_update()
 
-    def add_policy(self, brain_parameters: BrainParameters) -> None:
+    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
-            self.policy = MultiGpuPPOPolicy(
+            policy = MultiGpuPPOPolicy(
                 self.seed,
                 brain_parameters,
                 self.trainer_parameters,
         else:
-            self.policy = PPOPolicy(
+            policy = PPOPolicy(
                 self.seed,
                 brain_parameters,
                 self.trainer_parameters,
 
-        for _reward_signal in self.policy.reward_signals.keys():
+        for _reward_signal in policy.reward_signals.keys():
-    def get_policy(self, brain_name: str) -> PPOPolicy:
-        return self.policy
+        return policy
 
 
 def discount_rewards(r, gamma=0.99, value_next=0.0):

ml-agents/mlagents/trainers/rl_trainer.py

 from typing import Dict, List, Any, NamedTuple
 import numpy as np
 
-from mlagents.trainers.brain import BrainInfo
+from mlagents.trainers.brain import BrainParameters, BrainInfo
+from mlagents.trainers.tf_policy import TFPolicy
 from mlagents.trainers.components.reward_signals import RewardSignalResult
 
 LOGGER = logging.getLogger("mlagents.trainers")
         self.processing_buffer = ProcessingBuffer()
         self.update_buffer = AgentBuffer()
         self.episode_steps = {}
+        self.policy: TFPolicy = None
 
     def construct_curr_info(self, next_info: BrainInfo) -> BrainInfo:
         """
 
     def add_experiences(
         self,
+        name_behavior_id: str,
         curr_info: BrainInfo,
         next_info: BrainInfo,
         take_action_outputs: ActionInfoOutputs,
+        :param name_behavior_id: string policy identifier.
         :param curr_info: current BrainInfo.
         :param next_info: next BrainInfo.
         :param take_action_outputs: The outputs of the Policy's get_action method.
         raise UnityTrainerException(
             "The add_rewards_outputs method was not implemented."
         )
+
+    def add_policy(self, brain_parameters: BrainParameters) -> None:
+        """
+        Adds policy to trainers list of policies
+        """
+        policy = self.create_policy(brain_parameters)
+        self.policy = policy
+        self.policies[brain_parameters.brain_name] = policy

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

 from mlagents.envs.brain import BrainParameters, BrainInfo
 from mlagents.envs.action_info import ActionInfoOutputs
 from mlagents.envs.timers import timed
+from mlagents.trainers.tf_policy import TFPolicy
 from mlagents.trainers.sac.policy import SACPolicy
 from mlagents.trainers.rl_trainer import RLTrainer, AllRewardsOutput
 
         )
 
     def process_experiences(
-        self, current_info: BrainInfo, next_info: BrainInfo
+        self, name_behavior_id: str, current_info: BrainInfo, next_info: BrainInfo
+        :param name_behavior_id: string policy identifier.
         :param current_info: current BrainInfo.
         :param next_info: next BrainInfo.
         """
             self.update_reward_signals()
             self.trainer_metrics.end_policy_update()
 
-    def add_policy(self, brain_parameters: BrainParameters) -> None:
-        self.policy = SACPolicy(
+    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
+        policy = SACPolicy(
             self.seed,
             brain_parameters,
             self.trainer_parameters,
-        for _reward_signal in self.policy.reward_signals.keys():
+        for _reward_signal in policy.reward_signals.keys():
-    def get_policy(self, brain_name: str) -> SACPolicy:
-        return self.policy
+        return policy
 
     def update_sac_policy(self) -> None:
         """

ml-agents/mlagents/trainers/trainer.py

         self.summary_writer = tf.summary.FileWriter(self.summary_path)
         self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
         self.policy: TFPolicy = None
+        self.policies: Dict[str, TFPolicy] = {}
         self.step: int = 0
 
     def check_param_keys(self):
 
     def add_experiences(
         self,
+        name_behavior_id: str,
         curr_info: BrainInfo,
         next_info: BrainInfo,
         take_action_outputs: ActionInfoOutputs,
+        :param name_behavior_id: string policy identifier.
         :param curr_info: current BrainInfo.
         :param next_info: next BrainInfo.
         :param take_action_outputs: The outputs of the Policy's get_action method.
     def process_experiences(
-        self, current_info: BrainInfo, next_info: BrainInfo
+        self, name_behavior_id: str, current_info: BrainInfo, next_info: BrainInfo
+        :param name_behavior_id: string policy identifier.
         :param current_info: current BrainInfo.
         :param next_info: next BrainInfo.
         """
         """
         raise UnityTrainerException("The update_model method was not implemented.")
 
-    def add_policy(self, brain_parameters: BrainParameters) -> None:
+    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
-        Adds policy to trainers list of policies
+        Creates policy
-        Adds policy to trainers list of policies
+        Gets policy from trainers list of policies
-        raise UnityTrainerException("The update_model method was not implemented.")
+        return self.policies[brain_name]

ml-agents/mlagents/trainers/trainer_controller.py

                     brain_name, _ = name_behavior_id.split("?")
 
                     # This could be done with a try/except which may improve performance?
-                    if brain_name in self.trainers:
+                    try:
-                    else:
+                    except KeyError:
                         trainer = self.trainer_factory.generate(brain_name)
                         self.trainers[brain_name] = trainer
                         self.logger.info(trainer)
                 for behavior_identifier in self.brain_name_to_identifier[brain_name]:
                     if step_info.has_actions_for_brain(behavior_identifier):
                         trainer.add_experiences(
+                            behavior_identifier,
                             step_info.previous_all_brain_info[behavior_identifier],
                             step_info.current_all_brain_info[behavior_identifier],
                             step_info.brain_name_to_action_info[
                         trainer.process_experiences(
+                            behavior_identifier,
                             step_info.previous_all_brain_info[behavior_identifier],
                             step_info.current_all_brain_info[behavior_identifier],
                         )