SA_PGPolicy

doudizhu/apps/game/policy/SA_PGPolicy.py

@@ -0,0 +1,50 @@
+import sys
+from .PGPolicy import LearningPolicy, PGPolicy, rule
+
+'''
+Use summary act instead of actual act......
+(more) see SA_DQNPolicy.py
+
+'''
+
+class SA_PGPolicy(PGPolicy):
+
+    def call_score(self, state, default_action=None, learning=True):
+        '''
+        call a score from 1-3, if learning is true, use epsilon greedy
+        '''
+        vec = LearningPolicy._get_state_vec_sv2(state)
+        mask = self._get_call_score_mask_sa(state)
+        self.memory.set_state((vec, mask))
+        # epsilon-greedy
+        # generate action_idx and action
+        action_idx = self.model.choose(vec, mask, self._random_rand())
+        action = LearningPolicy.CALL_SCORE_ACTIONS[action_idx]
+        # record action
+        self.memory.set_action((action_idx, action))
+        # reward: turn punish - call is not involved so 0
+        self.memory.set_reward(0)
+        return action
+
+    def shot_poker(self, state, default_action=None, learning=True):
+        '''
+        shot a valid poker set, if learning is true, use epsilon greedy
+        '''
+        hand_pokers = state['hand_pokers']
+
+        vec = LearningPolicy._get_state_vec_sv2(state)
+        mask = self._get_shot_poker_mask_sa(state)
+        self.memory.set_state((vec, mask))
+        # generate action_idx and action
+        action_idx = self.model.choose(vec, mask, self._random_rand())
+        action = self._sa_idx_to_pokers(action_idx, state)
+        # record action
+        self.memory.set_action((action_idx, action))
+        # reward: turn punish
+        self.memory.set_reward(self.turn_reward)
+        return action
+
+    def __str__(self):
+        if self.comment is None:
+            self.comment = 'NOCOMMENT'
+        return "SA_PGPolicy-sv2[{}]-seed({}).{}.v{}".format(str(self.model),  self.seed, self.comment, self.generation)

doudizhu/simulate.py

@@ -12,15 +12,17 @@
 
 from apps.game.policy.DRL.REINFORCE_MLP import REINFORCE_MLP
 from apps.game.policy.PGPolicy import PGPolicy
+from apps.game.policy.SA_PGPolicy import SA_PGPolicy
 
 
-player = "Kate"
+player = "Ken"
 env = 'env1'
 epoch_num = 5000  
 display = False 
 save = True
 '''
 # test setting
+epoch_num = 1  
 display = True 
 save = False
 '''
@@ -106,6 +108,10 @@
     a1_model = DQNMLP(action_dim=DQNPolicy.S_ACTION_DIM, hidden_dims=(2048, )*8, learning_rate=1e-3)
     a1_policy = SA_DQNPolicy(a1_model, seed=0, comment="test_sa_"+env, e_greedy=(3e-1, -1e-4), save_every=1000)
     a1 = Agent(player, a1_policy)
+elif player == "Ken":
+    a1_model = REINFORCE_MLP(action_dim=DQNPolicy.S_ACTION_DIM, hidden_dims=(2048, )*8, learning_rate=1e-3)
+    a1_policy = SA_PGPolicy(a1_model, seed=0, comment="test_sa_"+env, save_every=1000)
+    a1 = Agent(player, a1_policy)
 else:
     raise NotImplementedError