diff --git a/RandomBird/FlappyAgent.py b/RandomBird/FlappyAgent.py
index 9f3ec84..24973e2 100644
--- a/RandomBird/FlappyAgent.py
+++ b/RandomBird/FlappyAgent.py
@@ -1,9 +1,66 @@
+from ple.games.flappybird import FlappyBird
+from ple import PLE
+
+from keras import optimizers
+from keras.models import Sequential
+from keras.layers import Dense, Flatten
+
 import numpy as np
 
+list_actions = [None,119]
+nb_save = 4
+save_pipe_center = []
+
+## Initialisation du réseau de neurones :
+batchSize = 256 # mini batch size
+## couches du réseau de neurone : plusieurs couches ne permettent pas de converger plus rapidement par expérience.
+dqn = Sequential()
+# 1st layer
+#dqn.add(Dense(units=112, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# 2nd layer
+dqn.add(Dense(units=500, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# 3rd layer
+#dqn.add(Dense(units=112, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# output layer
+dqn.add(Dense(units=2, init='lecun_uniform', activation="linear"))
+dqn.compile(loss="mean_squared_error", optimizer=optimizers.Adam(1e-4))
+
+dqn.load_weights("test_final.dqf") # Permet de charger le résultat précédent.
+
 def FlappyPolicy(state, screen):
-    action=None
-    if(np.random.randint(0,2)<1):
-        action=119
-    return action
+    return NemoPolicy(state)
+    
 
+def DoriPolicy(state):
+    next_pip_center = (state['next_pipe_bottom_y']+state['next_pipe_top_y'])/2
+    if state['player_y'] > next_pip_center:
+        return list_actions[1]
+    else:
+        return list_actions[0]
 
+def NemoPolicy(state):
+    global save_pipe_center, nb_save
+    
+    next_pipe_center = (state['next_pipe_bottom_y']+state['next_pipe_top_y'])/2
+    
+    if len(save_pipe_center)==0:
+        save_pipe_center = [next_pipe_center for i in range(nb_save+1)]
+    else:
+        save_pipe_center.append(next_pipe_center)
+        
+    if state['player_y'] > save_pipe_center.pop(0):
+        return list_actions[1]
+    else:
+        return list_actions[0]
+    
+    
+def greedy_action(network, state, batchSize): #Cherche la meilleure action prédite.
+    qval = network.predict(state.reshape(1,len(state)), batch_size=batchSize)
+    qval_av_action = [-9999]*2
+    for ac in range(0,2):
+        qval_av_action[ac] = qval[0][ac]
+    action = (np.argmax(qval_av_action))
+    return action
+    
+def Reinforcement_learning_policy(state):
+    return list_actions(greedy_action(dqn, state, batchSize))
\ No newline at end of file
diff --git a/RandomBird/Training.py b/RandomBird/Training.py
new file mode 100644
index 0000000..79ab54f
--- /dev/null
+++ b/RandomBird/Training.py
@@ -0,0 +1,157 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Mar  2 15:24:01 2018
+
+Fichier d'apprentissage du Flappy Bird
+
+@author: Gaspard Berthelin
+"""
+
+from ple.games.flappybird import FlappyBird
+from ple import PLE
+
+from keras import optimizers
+from keras.models import Sequential
+from keras.layers import Dense, Flatten
+
+import numpy as np
+
+## Initialisation du jeu : 
+game = FlappyBird(graphics="fixed")
+p = PLE(game, fps=30, frame_skip=1, num_steps=1, force_fps=True, display_screen=True)
+p.init()
+list_actions=[None,119]
+
+## Initialisation du réseau de neurones :
+batchSize = 256 # mini batch size
+## couches du réseau de neurone : plusieurs couches ne permettent pas de converger plus rapidement par expérience.
+dqn = Sequential()
+# 1st layer
+#dqn.add(Dense(units=112, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# 2nd layer
+dqn.add(Dense(units=500, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# 3rd layer
+#dqn.add(Dense(units=112, init='lecun_uniform', activation="relu", input_shape=(8,)))
+# output layer
+dqn.add(Dense(units=2, init='lecun_uniform', activation="linear"))
+dqn.compile(loss="mean_squared_error", optimizer=optimizers.Adam(1e-4))
+
+dqn.load_weights("final.dqf") # Permet de charger le résultat précédent.
+
+def process_state(state): #Adaptaion de l'état pour en faire une entrée du réseau de neurone.
+    return np.array(list(state.values()))
+
+def epsilon(step,total): # Définie la probabilité de choisir le cas aléatoire.
+    p_max = 0.4
+    p_min = 0.2
+    d = 0.0
+    x = step/total
+    if x < d:
+        return p_max
+    elif x > (1-d):
+        return p_min
+    return (p_min-p_max) * (x-d)/(1-2*d) + p_max
+
+def clip_reward(r): #Récompense sur l'évaluation des résultats (permet de donner le nombre de tuyaux passés)
+    rr=0
+    if r>0:
+        rr=1
+    if r<0:
+        rr=0
+    return train4000score13.0
+
+def training_reward(r): #Récompense du training : fort handicap sur le game over.
+    rr=0
+    if r>0:
+        rr=1
+    if r<0:
+        rr=-1000
+    return rr
+
+def random_action(state): # Pour accélérer la convergence, j'ai essayé un état aléatoire plus cohérent. Ce ne fut pas concluant à long terme.
+    if state[0] > (state[3]+state[4])/2.0 :
+        return 1
+    else :
+        return 0        
+    
+def greedy_action(network, state, batchSize): #Cherche la meilleure action prédite.
+    qval = network.predict(state.reshape(1,len(state)), batch_size=batchSize)
+    qval_av_action = [-9999]*2
+    for ac in range(0,2):
+        qval_av_action[ac] = qval[0][ac]
+    action = (np.argmax(qval_av_action))
+    return action
+
+def MCeval(network, trials): #Evalue la qualité du réseau jusqu'à présent.
+    scores = np.zeros((trials))
+    for i in range(trials):
+        p.reset_game()
+        while not(p.game_over()):
+            state = game.getGameState()
+            state = process_state(state)
+            action = greedy_action(network, state, batchSize)
+            action = list_actions[action]
+            reward = p.act(action)
+            reward = clip_reward(reward)
+            state = game.getGameState()
+            state = process_state(state)
+            scores[i] = scores[i] +  reward             
+    return np.sum(scores)
+
+
+## Training loop :
+total_games = 15000      #nombre de parties jouées pour l'entraînement.
+evaluation_period = 1000 #Tout les ... évalue la qualité du réseau.
+gamma = 0.99             #Permet de déifinir la récompense update.
+step_game = 0            #indice du nombre de parties jouées.
+while (step_game < total_games):
+    p.reset_game()       #réinitialisation du jeu
+    state = game.getGameState()
+    state = process_state(state)
+    rand_sum = 0
+    greedy_sum = 0
+    tuyau_passe = 0
+    while(not game.game_over()):
+        
+        if (np.random.random() < epsilon(step_game,total_games)): 
+            #Exploration
+            rand_sum = rand_sum + 1
+            #action = random_action(state)
+            action = np.random.choice([0,1])
+        else: 
+            #On suit le résultat du réseau de neurone.
+            greedy_sum = greedy_sum + 1
+            action = greedy_action(dqn, state, batchSize)
+        
+        #Résultat de l'action :
+        reward = p.act(list_actions[action])
+        reward = training_reward(reward)
+        if reward > 0:
+            tuyau_passe = tuyau_passe + 1
+        
+        new_state = game.getGameState()
+        new_state = process_state(new_state)
+        
+        terminal = game.game_over()   
+        
+        ## Update du réseau de neurone : 
+        newQ = dqn.predict(new_state.reshape(1,len(state)), batch_size=batchSize)
+        maxQ = np.max(newQ)
+        y = np.zeros((1,2))
+        y = dqn.predict(new_state.reshape(1,len(state)), batch_size=batchSize)
+        
+        update = reward + gamma * (1-terminal) * maxQ
+
+        y[0][action] = update
+        
+        dqn.fit(state.reshape(1, len(state)), y, batch_size=batchSize, nb_epoch=3, verbose=0)
+        state = new_state
+        
+    print("game %d : score = %d / prop_alea = %s pourcents" % (step_game ,tuyau_passe,round(rand_sum/(rand_sum+greedy_sum)*100)))
+    if step_game%evaluation_period == 0:
+        mcval = MCeval(dqn,50)
+        print('eval_MC=',MCeval(dqn,50))
+        dqn.save("train"  + str(step_game) + ".dqf")
+    
+    step_game = step_game + 1
+dqn.save("final.dqf")       
\ No newline at end of file
diff --git a/RandomBird/test_final.dqf b/RandomBird/test_final.dqf
new file mode 100644
index 0000000..382f62a
Binary files /dev/null and b/RandomBird/test_final.dqf differ
diff --git a/RandomBird/train0.dqf b/RandomBird/train0.dqf
new file mode 100644
index 0000000..9ca4bd4
Binary files /dev/null and b/RandomBird/train0.dqf differ
diff --git a/RandomBird/train1000.dqf b/RandomBird/train1000.dqf
new file mode 100644
index 0000000..22f2641
Binary files /dev/null and b/RandomBird/train1000.dqf differ
diff --git a/RandomBird/train10000.dqf b/RandomBird/train10000.dqf
new file mode 100644
index 0000000..a4401f5
Binary files /dev/null and b/RandomBird/train10000.dqf differ
diff --git a/RandomBird/train11000.dqf b/RandomBird/train11000.dqf
new file mode 100644
index 0000000..d803871
Binary files /dev/null and b/RandomBird/train11000.dqf differ
diff --git a/RandomBird/train12000.dqf b/RandomBird/train12000.dqf
new file mode 100644
index 0000000..9e5701c
Binary files /dev/null and b/RandomBird/train12000.dqf differ
diff --git a/RandomBird/train13000.dqf b/RandomBird/train13000.dqf
new file mode 100644
index 0000000..95c8d9a
Binary files /dev/null and b/RandomBird/train13000.dqf differ
diff --git a/RandomBird/train2000.dqf b/RandomBird/train2000.dqf
new file mode 100644
index 0000000..30258ac
Binary files /dev/null and b/RandomBird/train2000.dqf differ
diff --git a/RandomBird/train3000.dqf b/RandomBird/train3000.dqf
new file mode 100644
index 0000000..a78453e
Binary files /dev/null and b/RandomBird/train3000.dqf differ
diff --git a/RandomBird/train4000.dqf b/RandomBird/train4000.dqf
new file mode 100644
index 0000000..909223b
Binary files /dev/null and b/RandomBird/train4000.dqf differ
diff --git a/RandomBird/train5000.dqf b/RandomBird/train5000.dqf
new file mode 100644
index 0000000..2679ffb
Binary files /dev/null and b/RandomBird/train5000.dqf differ
diff --git a/RandomBird/train6000.dqf b/RandomBird/train6000.dqf
new file mode 100644
index 0000000..0b41f07
Binary files /dev/null and b/RandomBird/train6000.dqf differ
diff --git a/RandomBird/train7000.dqf b/RandomBird/train7000.dqf
new file mode 100644
index 0000000..bc79deb
Binary files /dev/null and b/RandomBird/train7000.dqf differ
diff --git a/RandomBird/train8000.dqf b/RandomBird/train8000.dqf
new file mode 100644
index 0000000..fa2e09d
Binary files /dev/null and b/RandomBird/train8000.dqf differ
diff --git a/RandomBird/train9000.dqf b/RandomBird/train9000.dqf
new file mode 100644
index 0000000..00392e1
Binary files /dev/null and b/RandomBird/train9000.dqf differ