def main(lr=0.001, episodeMemory=100, replaySize=64, gamma=0.95):
np.random.seed(0)
env = gym.make('MountainCar-v0')
model = keras.Sequential()
model.add(
Dense(128, activation="relu", input_dim=3,
kernel_initializer='normal'))
model.add(Dense(52, activation="relu"))
model.add(Dense(1, kernel_initializer='normal', activation="linear"))
adam = keras.optimizers.Adam(lr=lr)
model.compile(loss='mean_squared_error', optimizer=adam)
#gamma = 0.95
memorySize = 200 * episodeMemory
dqn = DQN(model, gamma, memorySize, replaysize=replaySize, _env=env)
dqnScore = dqnScorerMountainCar(dqn, _env=env)
nrofEpisodes = 1001
#nrofEpisodes = 20
res = np.zeros(shape=(nrofEpisodes, 2))
for episode in range(nrofEpisodes):
env.reset()
action = 0
obs, _, done, _ = env.step(action)
#if (episode % 100) == 10:
if (episode % 100) == 10:
print("episode ", episode)
dqnScore.printDistance()
#dqnScore.plot_cost_to_ßgo_mountain_car()
#print(res[episode-1,:])
print("--- %s seconds ---" % (time.time() - start_time))
iter = 0
while not done:
iter += 1
action = dqn.action(obs)
new_obs, reward, done, info = env.step(action)
if (done and (iter < 199)):
reward = (200 - iter) / 10
print("****Success*****", -iter)
dqn.add(action, obs, new_obs, reward)
obs = new_obs
#if(episode % 100) == 10:
# env.render()j
dqn.replay()
env.reset()
dqnScore.updateResult(iter)
#res[episode,:] = [np.min(x[:,0]),np.max(x[:,0])]
title = "eps_%d_mem_%d_rep_%d_gamma_%d" % (nrofEpisodes, episodeMemory,
replaySize, gamma * 100)
dqnScore.plotResults(title)
dqnScore.plot_cost_to_go_mountain_car(title)
Q = model(obs, save=False)
#print Q[0]
if epsilon > rand() or step < 100:
action = randint(0, 2)
else:
action = np.argmax(Q[0])
epsilon -= 2e-4
if epsilon < 0.:
epsilon = 0.
obs, reward, done, _ = env.step(action)
reward = 0.
if done:
reward = -1.
episode += 1
Memory.add(last_obs, action, reward, obs, done)
if done:
obs = env.reset()
last_obs = deepcopy(obs)
if done and episode % 100 == 0:
print 'episode:', episode, 'step:', step, 'eps:', epsilon, 'ave:', time / 100., 'Q:', Q[
0]
time = 0.
#t = deepcopy(Q)
if step < 100:
continue
sample = [Memory.ReplayMemory[(Memory.count - 1) % 10**6]] #sample(16)
#sample = []
#for i in range(10):