cfr:test

README.md

@@ -0,0 +1,16 @@
+
+start:cfr([JQ],"",1,1)
+
++ strat_0:J.[0.7,0.3]
+0:cfr([JQ],"0",0.7,1)
++ strat_1:Q.0[0.6,0.4]
+    0:cfr([JQ],"00",0.7,0.6) -> terminal,get_uitl
++ strat_0:J.01[0.8,0.2]    
+    1:cfr([JQ],"01",0.7,0.4)
+        0:cfr([JQ],"010",0.56,0.4) -> t
+        1:cfr([JQ],"011",0.14,0.4) -> t
+
+1:cfr([JQ],"1",0.3,1)
++ strat_1:Q.1[0.5,0.5]
+    0:cfr([JQ],"10",0.3,0.5) -> t
+    1:cfr([JQ],"11",0.3,0.5) -> t

src/cfr/calcu.py

@@ -9,65 +9,116 @@ game = KuhnPoker()
 def cfr(cards, his, p0, p1):
 
     if game.is_terminal(his):
-        return game.get_profit(cards, his, 0)
+        return game.get_util(cards, his, 0)
     
     player = game.get_cur_player(his)
     
-    # + player ?
-    info_key = game.get_Info_set(cards[player], his, player)
+    info_key = game.get_Info_set(cards[player], his)
     
     if info_key not in info_map:
         info_map[info_key] = InfoSet(2) # 0-1
     info = info_map[info_key]
     
-    wgt = p0 if player == 0 else p1
-    strat = info.get_strat(wgt)
+    reach_prob = p0 if player == 0 else p1
+    strat = info.get_strat(reach_prob)
     
 
-    act_profit = [0.0, 0.0]
-    info_profit = 0.0 # 期望
+    act_util = [0.0, 0.0]
+    info_util = 0.0
     
     for action in [0, 1]:
         next_his = his + str(action)
         
         if player == 0:
-            profit = cfr(cards, next_his, p0*strat[action], p1)
+            util = cfr(cards, next_his, p0*strat[action], p1)
         else:
-            profit = cfr(cards, next_his, p0, p1*strat[action])
+            util = cfr(cards, next_his, p0, p1*strat[action])
         
-        act_profit[action] = profit
-        info_profit += strat[action]*profit
+        act_util[action] = util
+        info_util += strat[action]*util
 
-    # 更新
     for action in [0, 1]:
+        p0_util = act_util[action] - info_util
         if player == 0:
-            regret = act_profit[action] - info_profit
+            regret = p0_util
             other_r = p1
         else:
-            regret = -(act_profit[action] - info_profit)
+            regret = -p0_util
             other_r = p0
+        # 对手的reach probe加权累计regret
+        # 假设对手到达这的概率小，那么这个regret就不重要
         info.regret_sum[action] += other_r * regret
     
-    return info_profit
+    return info_util
 
 
-def test():
+def test(cnt):
     
     cards = ['J', 'Q', 'K']
-    p_sum = 0.0
+    util_sum = 0.0
     
-    for i in range(10):
+    for i in range(cnt):
         card_r = random.sample(cards, 2)
+        util = cfr(card_r, "", 1.0, 1.0)
+        util_sum += util
         
-        pf = cfr(card_r, "", 1.0, 1.0)
-        p_sum += pf
+        if (i + 1) % cnt == 0:
+            util_avg = util_sum / (i + 1)
+            sum_regret = 0
+            for info in info_map.values():
+                max_r = max(info.regret_sum) if info.regret_sum else 0
+                sum_regret += max_r
+            avg_regret = sum_regret / (i + 1)
+            print(f"range {i+1}/{cnt}, rvg_regret: {avg_regret:.3f}, avg_util: {util_avg:.3f}")
+    final_util = util_sum / cnt
+    final_regret = sum(max(0, max(info.regret_sum)) for info in info_map.values()) / cnt
+    print(f"final_avg_regret: {final_regret:.3f}, final_avg_util: {final_util:.3f}")
+    return util_sum / cnt
+
+def all_avg_strat():
+    avg_strat = {}
+    for info_key, info in info_map.items():
+        avg_strat[info_key] = info.get_avg_strat()
+    return avg_strat
+
+def print_strat(strat):
+    if strat is None:
+        strat = all_avg_strat()
+
+    print("="*60)
+    print(f"{'Info Set':<15} {'Check/Fold':<15} {'Bet/Call':<15}")
+    print("-"*60)
+    for info_key in sorted(strat.keys()):
+        probe = strat[info_key]
+        parts = info_key.split('.')
+        card = parts[0]
+        his = parts[1] if len(parts) > 1 else ''
+
+        act0 = game.get_act_name(his, 0)
+        act1 = game.get_act_name(his, 1)
+
+        print(f"{info_key:<15} {act0}:{probe[0]:<10.3f} {act1}:{probe[1]:<10.3f}")
+    print("="*60)
+
+def reset():
+    global info_map
+    info_map = {}
+
+
+
+if __name__ == "__main__":
+    reset()
+    test(100000)
+    print("平均策略")
+    print_strat(None)
+
+    print("策略check")
+    strats = all_avg_strat()
+    print(f"  K_Bet_Probe:    {strats.get('K.', [0, 0])[1]:.3f} ")
+    print(f"  Q_Check_Probe:  {strats.get('Q.', [0, 0])[0]:.3f} ")
+    print(f"  J_Check_Probe:    {strats.get('J.', [0, 0])[0]:.3f} ")
+
         
-        if (i + 1) % 10 == 0:
-            avg_p = p_sum / (i + 1)
-            print(f"Range {i+1}/10, Avg : {avg_p:.3f}")
-    
-    return p_sum / 10
 
-def print_strat():
-    
 
+    

src/cfr/info_set.py

@@ -2,35 +2,35 @@ class InfoSet:
     def __init__(self, act_cnt=2): # 0-1
         self.act_cnt = act_cnt
         
-        self.regret_sum = [0.0] * act_cnt
-        self.strat = [0.0] * act_cnt        
-        self.strat_sum = [0.0] * act_cnt    
+        self.regret_sum = [0.0] * act_cnt  # 累积regret，用于更新策略
+        self.strat = [0.0] * act_cnt       
+        self.strat_sum = [0.0] * act_cnt   # 累积策略, 用于平均策略
         
-    def get_strat(self, wgt):
+    # 更新策略 论文Part3(8)
+    def get_strat(self, reach_prob):
 
         normal = 0.0
         for i in range(self.act_cnt):
-            if self.regret_sum[i] > 0:
-                self.strat[i] = self.regret_sum[i]
-            else:
-                self.strat[i] = 0.0
+            self.strat[i] = max(self.regret_sum[i], 0.0)
             normal += self.strat[i]
         
         if normal > 0:
             for i in range(self.act_cnt):
                 self.strat[i] /= normal
         else:
-            ##
+            # 混合策略
             prob = 1.0/self.act_cnt
             for i in range(self.act_cnt):
                 self.strat[i] = prob
-            # return
-        
+                
+        # 更新累计策略（拆开）
+        # 加权累计
         for i in range(self.act_cnt):
-            self.strat_sum[i] += wgt * self.strat[i]
+            self.strat_sum[i] += reach_prob * self.strat[i]
             
         return self.strat
     
+    # 论文Part2(4)
     def get_avg_strat(self):
         avg_strat = [0.0] * self.act_cnt
         normal = sum(self.strat_sum)
@@ -39,10 +39,8 @@ class InfoSet:
             for i in range(self.act_cnt):
                 avg_strat[i] = self.strat_sum[i] / normal
         else:
-            ##
             prob = 1.0/self.act_cnt
             for i in range(self.act_cnt):
-                avg_strat[i] = prob
-            # return            
+                avg_strat[i] = prob            
         return avg_strat
     

src/game/kuhn_poker.py

@@ -3,11 +3,10 @@ class KuhnPoker:
         self.cards = ['J', 'Q', 'K']
         self.actions = [0, 1]  # Check/Fold=0, Bet/Call=1
     
-    # 结构是不是要修改下？
     def is_terminal(self, history):
         return history in ['00', '10', '010', '011', '11']
     
-    def get_profit(self, cards, history, player):
+    def get_util(self, cards, history, player):
         if not self.is_terminal(history):
             return 0.0
             
@@ -42,7 +41,7 @@ class KuhnPoker:
                 
         return 0.0
     
-    def get_Info_set(self, card, history, player):
+    def get_Info_set(self, card, history):
         return f"{card}.{history}"
     
     def get_cur_player(self, history):

tests/test_game.py

@@ -22,7 +22,7 @@ def test_game():
         (['Q', 'K'], '011', 1),
     ]
     for cards, his, player in test_cases:
-        profit = game.get_profit(cards, his, player)
+        profit = game.get_util(cards, his, player)
         print(f"player{player}profit: {profit}")
     
     info_cases = [