Initial commit

Divers/renforcement2/CartPole_train.py

@@ -0,0 +1,82 @@
+import gym
+import numpy as np
+import cv2
+import CartPole_common
+
+env=gym.make("CartPole-v0")
+env._max_episode_steps=500
+
+alpha=0.05
+gamma=0.98
+
+epoch=50000
+show_every=500
+
+epsilon=1.
+epsilon_min=0.05
+start_epsilon=1
+end_epsilon=epoch//2
+epsilon_decay_value=epsilon/(end_epsilon-start_epsilon)
+
+nbr_action=env.action_space.n
+q_table=np.random.uniform(low=-1, high=1, size=(CartPole_common.division+[nbr_action]))
+
+result_done=0
+scores=[]
+best_score=0
+for episode in range(epoch):
+    obs=env.reset()
+    discrete_state=CartPole_common.discretise(obs)
+    done=False
+
+    if episode%show_every == 0:
+        render=True
+        mean_score=np.mean(scores)
+        print("Epoch {:06d}/{:06d} reussite:{:04d}/{:04d} epsilon={:06.4f} Mean score={:08.4f} alpha={:06.4f}".format(episode, epoch, result_done, show_every, epsilon, mean_score, alpha))
+        scores=[]
+        result_done=0
+        if mean_score>best_score:
+            print("Sauvegarde ...")
+            np.save("CartPole_qtable", q_table)
+            best_score=mean_score
+        alpha=alpha*0.99
+
+    else:
+        render=False
+
+    score=1
+    while not done:
+
+        if np.random.random()>epsilon:
+            action=np.argmax(q_table[discrete_state])
+        else:
+            action=np.random.randint(nbr_action)
+
+        new_state, reward, done, info=env.step(action)
+        new_discrete_state=CartPole_common.discretise(new_state)
+
+        if episode%show_every == 0:
+            env.render()
+
+        #reward=2-np.abs(new_state[0])
+        if done:
+            scores.append(score)
+            if score==env._max_episode_steps:
+                result_done+=1
+            else:
+                reward=-10
+
+        max_future_q=np.max(q_table[new_discrete_state])
+        current_q=q_table[discrete_state][action]
+        new_q=(1-alpha)*current_q+alpha*(reward+gamma*max_future_q)
+        q_table[discrete_state][action]=new_q
+
+        score+=1
+        discrete_state=new_discrete_state
+
+    if end_epsilon>=episode>=start_epsilon:
+        epsilon-=epsilon_decay_value
+        if epsilon<epsilon_min:
+            epsilon=epsilon_min
+
+env.close()