# serialize model to JSON
model_save = model.to_json()
with open("save/NNmodel1.json", "w") as json_file:
json_file.write(model_save)
print("Saved model to disk!")
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=10000, window_length=1)
#policy1 = BoltzmannQPolicy()
policy1 = EpsDisGreedyQPolicy(eps=0.0001, eps_decay=0.999)
policy2 = BoltzmannQPolicy()
callback12 = FileLogger(filepath='save/history12_{}'.format(timenow),
interval=1)
dqn8 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
gamma=0.99,
nb_steps_warmup=2000,
target_model_update=1,
policy=policy1)
dqn8.compile(Adam(lr=1e-3), metrics=['mae'])
history8 = dqn8.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback12],
verbose=2)
callback1 = FileLogger(filepath='save/nhistory1_{}'.format(timenow), interval=1)
callback2 = FileLogger(filepath='save/nhistory2_{}'.format(timenow), interval=1)
callback3 = FileLogger(filepath='save/nhistory3_{}'.format(timenow), interval=1)
callback4 = FileLogger(filepath='save/nhistory4_{}'.format(timenow), interval=1)
callback5 = FileLogger(filepath='save/nhistory5_{}'.format(timenow), interval=1)
# dqn1 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=1000,
# target_model_update=1e-2, policy=policy1)
# dqn1.compile(Adam(lr=1e-3), metrics=['mae'])
# history1 = dqn1.fit(env, nb_epsteps=5000, visualize=False, callbacks=[callback1], verbose=2)
dqn2 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, batch_size=32, nb_steps_warmup=1000,
target_model_update=1e-2, policy=policy2)
dqn2.compile(Adam(lr=0.01), metrics=['mse'])
# dqn2.save_weights('save/dqn_blotzmann0.8_{}_weights.h5f'.format(ENV_NAME), overwrite=True)
history2 = dqn2.fit(env, nb_steps=200000, visualize=False, callbacks=[callback1], verbose=2)
time.sleep(3600)
# dqn3 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=100,
# target_model_update=1e-2, policy=policy1, enable_double_dqn=False)
# dqn3.compile(Adam(lr=1e-3), metrics=['mae'])
# history3 = dqn3.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback3], verbose=2)
# dqn4 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=100,
# target_model_update=1e-2, policy=policy2, enable_double_dqn=False)
# dqn4.compile(Adam(lr=1e-3), metrics=['mae'])
# history4 = dqn4.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback4], verbose=2)
# print(history1.history.keys())
# print(len(history1.history['policy_config']))
#dqn1 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, batch_size=64, nb_steps_warmup=2000,
# target_model_update=1e-2, policy=policy2)
#dqn1.compile(Adam(lr=1e-3), metrics=['mae'])
#dqn1.fit(env, nb_steps=300000, visualize=False, callbacks=[callback1], verbose=2)
#dqn2 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, batch_size=64, gamma=0.5, nb_steps_warmup=2000,
# target_model_update=1e-2, policy=policy2)
#dqn2.compile(Adam(lr=1e-3), metrics=['mae'])
#dqn2.fit(env, nb_steps=300000, visualize=False, callbacks=[callback2], verbose=2)
dqn3 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, batch_size=64, nb_steps_warmup=2000,
target_model_update=1e-2, policy=policy1)
dqn3.compile(Adam(lr=1e-3), metrics=['mae'])
dqn3.fit(env, nb_steps=300000, visualize=False, callbacks=[callback3], verbose=2)
'''
pyplot.figure()
pyplot.subplot(2, 1, 1)
pyplot.plot(history1.history['episode_reward'], 'r--',history3.history['episode_reward'], 'b--')
pyplot.subplot(2, 1, 2)
#pyplot.plot(history1.history['nb_steps'], history1.history['episode_reward'], 'r', history2.history['nb_steps'], history2.history['episode_reward'], 'g')
pyplot.plot(history2.history['episode_reward'], 'r', history4.history['episode_reward'], 'b')
pyplot.show()
#pyplot.savefig('save/BoltzmannQPolicy')
'''
dqn1 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=1000,
target_model_update=1e-2, policy=policy1)
dqn1.compile(Adam(lr=1e-3), metrics=['mae'])
history1 = dqn1.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback1], verbose=2)
dqn2 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=1000,
target_model_update=1e-2, policy=policy2)
dqn2.compile(Adam(lr=1e-3), metrics=['mae'])
history2 = dqn2.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback2], verbose=2)
dqn3 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=1000,
target_model_update=1e-2, policy=policy1, enable_double_dqn=False)
dqn3.compile(Adam(lr=1e-3), metrics=['mae'])
history3 = dqn3.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback3], verbose=2)
'''
dqn3 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
batch_size=640,
nb_steps_warmup=20000,
target_model_update=1e-2,
policy=policy1)
dqn3.compile(Adam(lr=1e-2), metrics=['mae'])
history3 = dqn3.fit(env,
nb_steps=500000,
visualize=False,
callbacks=[callback1],
verbose=2)
#dqn3.save_weights('save/dqn4_{}_weights.h5f'.format(ENV_NAME), overwrite=True)
# serialize model to JSON
model_save = model.to_json()
with open("save/NNmodel1.json", "w") as json_file:
json_file.write(model_save)
print("Saved model to disk!")
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=10000, window_length=1)
#policy1 = BoltzmannQPolicy()
policy1 = EpsDisGreedyQPolicy(eps=0.0001, eps_decay=0.999)
policy2 = BoltzmannQPolicy()
callback10 = FileLogger(filepath='save/history10_{}'.format(timenow),
interval=1)
dqn10 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
gamma=0.99,
nb_steps_warmup=2000,
target_model_update=1e-2,
policy=policy2)
dqn10.compile(Adam(lr=1e-3), metrics=['mae'])
history10 = dqn10.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback10],
verbose=2)
#dqn3 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=100,
# target_model_update=1e-2, policy=policy1, enable_double_dqn=False)
#dqn3.compile(Adam(lr=1e-3), metrics=['mae'])
#history3 = dqn3.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback3], verbose=2)
dqn4 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
batch_size=64,
nb_steps_warmup=2000,
target_model_update=1e-2,
policy=policy2)
dqn4.compile(Adam(lr=1e-3), metrics=['mae'])
history4 = dqn4.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback4],
verbose=2)
'''
pyplot.figure()
pyplot.subplot(2, 1, 1)
pyplot.plot(history1.history['episode_reward'], 'r--',history3.history['episode_reward'], 'b--')
pyplot.subplot(2, 1, 2)
#pyplot.plot(history1.history['nb_steps'], history1.history['episode_reward'], 'r', history2.history['nb_steps'], history2.history['episode_reward'], 'g')
pyplot.plot(history2.history['episode_reward'], 'r', history4.history['episode_reward'], 'b')
pyplot.show()
#pyplot.savefig('save/BoltzmannQPolicy')
'''
model.add(Activation('linear'))
print(model.summary())
# serialize model to JSON
model_save = model.to_json()
with open("save/NNmodel1.json", "w") as json_file:
json_file.write(model_save)
print("Saved model to disk!")
memory = SequentialMemory(limit=10000, window_length=1)
#policy1 = BoltzmannQPolicy()
policy1 = EpsDisGreedyQPolicy(eps=0.0001, eps_decay=0.999)
policy2 = BoltzmannQPolicy()
callback6 = FileLogger(filepath='save/history6_{}'.format(timenow), interval=1)
dqn6 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
gamma=0.1,
nb_steps_warmup=2000,
target_model_update=1e-2,
policy=policy2)
dqn6.compile(Adam(lr=1e-3), metrics=['mae'])
history6 = dqn6.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback6],
verbose=2)
#dqn4 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=100,
# target_model_update=1e-2, policy=policy2, enable_double_dqn=False)
#dqn4.compile(Adam(lr=1e-3), metrics=['mae'])
#history4 = dqn4.fit(env, nb_epsteps=100, visualize=False, callbacks=[callback4], verbose=2)
dqn5 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
batch_size=96,
nb_steps_warmup=2000,
target_model_update=1e-2,
policy=policy2)
dqn5.compile(Adam(lr=1e-3), metrics=['mae'])
history5 = dqn5.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback5],
verbose=2)
'''
pyplot.figure()
pyplot.subplot(2, 1, 1)
pyplot.plot(history1.history['episode_reward'], 'r--',history3.history['episode_reward'], 'b--')
pyplot.subplot(2, 1, 2)
#pyplot.plot(history1.history['nb_steps'], history1.history['episode_reward'], 'r', history2.history['nb_steps'], history2.history['episode_reward'], 'g')
pyplot.plot(history2.history['episode_reward'], 'r', history4.history['episode_reward'], 'b')
pyplot.show()
#pyplot.savefig('save/BoltzmannQPolicy')
'''
model.add(Activation('relu'))
model.add(Dense(nb_actions))
model.add(Activation('linear'))
print(model.summary())
# serialize model to JSON
model_save = model.to_json()
with open("save/NNmodel1.json", "w") as json_file:
json_file.write(model_save)
print("Saved model to disk!")
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=10000, window_length=1)
#policy1 = BoltzmannQPolicy()
policy1 = EpsDisGreedyQPolicy(eps=0.0001, eps_decay=0.999)
policy2 = BoltzmannQPolicy()
callback9 = FileLogger(filepath='save/history9_{}'.format(timenow), interval=1)
dqn9 = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, gamma=0.99, nb_steps_warmup=2000,
target_model_update=1e-2, policy=policy1)
dqn9.compile(Adam(lr=1e-3), metrics=['mae'])
history9 = dqn9.fit(env, nb_epsteps=3000, visualize=False, callbacks=[callback9], verbose=2)
model.add(Activation('linear'))
print(model.summary())
# serialize model to JSON
model_save = model.to_json()
with open("save/NNmodel1.json", "w") as json_file:
json_file.write(model_save)
print("Saved model to disk!")
memory = SequentialMemory(limit=10000, window_length=1)
#policy1 = BoltzmannQPolicy()
policy1 = EpsDisGreedyQPolicy(eps=0.0001, eps_decay=0.999)
policy2 = BoltzmannQPolicy()
callback7 = FileLogger(filepath='save/history7_{}'.format(timenow), interval=1)
dqn7 = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
gamma=0.5,
nb_steps_warmup=2000,
target_model_update=1e-2,
policy=policy2)
dqn7.compile(Adam(lr=1e-3), metrics=['mae'])
history7 = dqn7.fit(env,
nb_epsteps=3000,
visualize=False,
callbacks=[callback7],
verbose=2)