refactoring

[shutbox.git] / src / qtable.py

diff --git a/src/qtable.py b/src/qtable.py
index 5dda44424da840faab10be2ac05307bcfaf33780..29a3081235bfa8675657cdc6d0b866158ee88f7a 100755 (executable)
--- a/src/qtable.py
+++ b/src/qtable.py
@@ -6,11 +6,11 @@ import sys
 
 from game import Game
 
-learning_rate = 0.1
+learning_rate = 0.001
 discount_factor = 1.0
 
-states_dim = 147456 # 2^12 * 6^2
-actions_dim = 637 # 12+1 * (6+1)^2
+states_dim = 36864 # 2^10 * 6^2
+actions_dim = 539 # 10+1 * (6+1)^2
 num_episodes = 10000000000
 
 def find_state_qid(shutable, diced):
@@ -18,13 +18,13 @@ def find_state_qid(shutable, diced):
     for rod in shutable:
         qid += pow(2, rod-1)
     for i in range(len(diced)):
-        qid += (diced[i]-1) * pow(6, i) * pow(2, 12)
+        qid += (diced[i]-1) * pow(6, i) * pow(2, 10)
     return qid
 
 def find_option_qid(option):
     qid = 0
     for i in range(len(option)):
-        qid += option[i] * pow(7, i) * pow(13, len(option)-1)
+        qid += option[i] * pow(7, i) * pow(11, len(option)-1)
     return qid
 
 def select_option(opts, qs):
@@ -34,14 +34,12 @@ def select_option(opts, qs):
         opt_qid = find_option_qid(opt)
         opt_qid_pairs.append( [opt, opt_qid] )
         opt_qsum += qs[opt_qid]
-    random.shuffle(opt_qid_pairs)
     ran_pt = random.uniform(0.0, opt_qsum)
     decision_pt = 0.0
     for opt_qid_pair in opt_qid_pairs:
         decision_pt += qs[ opt_qid_pair[1] ]
         if ran_pt <= decision_pt:
             return (opt_qid_pair[0], opt_qid_pair[1])
-    return (None, None)
 
 Q = np.ones([states_dim, actions_dim])
 
@@ -52,20 +50,20 @@ for i in range(num_episodes):
     g.dice()
     state_qid = find_state_qid(g.get_shutable(), g.get_diced())
     while not g.is_over():
-        opt, opt_qid = select_option( g.get_options(), Q[state_qid, :] )
-        if opt:
+        options = g.get_options()
+        if len(options) > 0:
+            opt, opt_qid = select_option( options, Q[state_qid, :] )
             old_score = g.get_score()
             g.shut(opt)
             g.dice()
-            reward = (g.get_score() - old_score) / 12.0
+            reward = g.get_score() - old_score
             new_state_qid = find_state_qid(g.get_shutable(), g.get_diced())
             Q[state_qid, opt_qid] += \
                 learning_rate * (reward
                                  + discount_factor * np.max(Q[new_state_qid, :])
                                  - Q[state_qid, opt_qid])
             state_qid = new_state_qid
-        else:
-            Q[state_qid, opt_qid] = 0
+    Q[state_qid, :] = 0.0
     running_score[0] *= 0.99999999
     running_score[0] += g.get_score()
     running_score[1] *= 0.99999999