我试图由著名的安德烈·卡帕蒂（Andrej Karpathy）实现完全相同的python编码，以训练RL特工打乒乓球 （https://gist.github.com/karpathy/a4166c7fe253700972fcbc77e4ea32c5），除了我将环境从Gym迁移到Retro，从Python2迁移到Python3。除了Retro中的动作空间是按索引而不是像Gym中那样离散的以外，其他所有内容都是相同的。索引的大小为8，索引4和5是向上和向下移动的动作。
但是，为什么稍加修改就导致特工在超过3,000次情节后根本无法以-20的运行奖励获得学习？
在对帧进行预处理之后，我检查了图像，然后将其输入到策略前向神经网络，这似乎是正常的。
我怀疑在神经网络的策略转发之后问题仍然存在，但是我只是无法找到问题所在。欣赏是否有人可以对其进行测试。
整个编码如下：
import retro
import numpy as np
import _pickle as pickle
H = 200 # number of hidden layer neurons
batch_size = 10 # every how many episodes to do a param update?
learning_rate = 1e-4
gamma = 0.98 # discount factor for reward
decay_rate = 0.98 # decay factor for RMSProp leaky sum of grad^2
resume = False # resume from previous checkpoint?
render = True
# model initialization
D = 80 * 80 # input dimensionality: 80x80 grid
if resume:
model = pickle.load(open('save.p', 'rb'))
else:
model = {}
model['W1'] = np.random.randn(H,D) / np.sqrt(D) # "Xavier" initialization
model['W2'] = np.random.randn(H) / np.sqrt(H)
grad_buffer = { k : np.zeros_like(v) for k,v in model.items() } # update buffers that add up gradients over a batch
rmsprop_cache = { k : np.zeros_like(v) for k,v in model.items() } # rmsprop memory
def sigmoid(x):
return 1.0 / (1.0 + np.exp(-x)) # sigmoid "squashing" function to interval [0,1]
def prepro(I):
""" prepro 210x160x3 uint8 frame into 6400 (80x80) 1D float vector """
I = I[35:195] # crop
I = I[::2,::2,0] # downsample by factor of 2
I[I == 144] = 0 # erase background (background type 1)
I[I == 109] = 0 # erase background (background type 2)
I[I != 0] = 1 # everything else (paddles, ball) just set to 1
return I.astype(np.float).ravel()
def discount_rewards(r):
""" take 1D float array of rewards and compute discounted reward """
discounted_r = np.zeros_like(r)
running_add = 0
for t in reversed(range(0, r.size)):
if r[t] != 0: running_add = 0 # reset the sum, since this was a game boundary (pong specific!)
running_add = running_add * gamma + r[t]
discounted_r[t] = running_add
return discounted_r
def policy_forward(x):
h = np.dot(model['W1'], x)
h[h<0] = 0 # ReLU nonlinearity
logp = np.dot(model['W2'], h)
p = sigmoid(logp)
return p, h # return probability of taking action 2, and hidden state
def policy_backward(eph, epdlogp, epx):
""" backward pass. (eph is array of intermediate hidden states) """
dW2 = np.dot(eph.T, epdlogp).ravel()
dh = np.outer(epdlogp, model['W2'])
dh[eph <= 0] = 0 # backpro prelu
dW1 = np.dot(dh.T, epx)
return {'W1':dW1, 'W2':dW2}
env=retro.make(game='Pong-Atari2600',players=1)
observation = env.reset()
prev_x = None # used in computing the difference frame
xs,hs,dlogps,drs = [],[],[],[]
running_reward = None
reward_sum = 0
episode_number = 0
while True:
if render: env.render()
action=[0,0,0,0,0,0,0,0] #reset the rl action
# preprocess the observation, set input to network to be difference image
cur_x = prepro(observation)
x = cur_x - prev_x if prev_x is not None else np.zeros(D)
prev_x = cur_x
# forward the policy network and sample an action from the returned probability
aprob, h = policy_forward(x)
rlaction = 4 if np.random.uniform() < aprob else 5 # roll the dice!
# record various intermediates (needed later for backprop)
xs.append(x) # observation
hs.append(h) # hidden state
y = 1 if rlaction == 4 else 0 # a "fake label"
dlogps.append(y - aprob) # grad that encourages the action that was taken to be taken (see http://cs231n.github.io/neural-networks-2/#losses if confused)
action[rlaction]=1
# step the environment and get new measurements
observation, reward, done, info = env.step(action)
reward_sum += reward
drs.append(reward) # record reward (has to be done after we call step() to get reward for previous action)
if done: # an episode finished
episode_number += 1
# stack together all inputs, hidden states, action gradients, and rewards for this episode
epx = np.vstack(xs)
eph = np.vstack(hs)
epdlogp = np.vstack(dlogps)
epr = np.vstack(drs)
xs,hs,dlogps,drs = [],[],[],[] # reset array memory
# compute the discounted reward backwards through time
discounted_epr = discount_rewards(epr)
# standardize the rewards to be unit normal (helps control the gradient estimator variance)
discounted_epr -= np.mean(discounted_epr)
discounted_epr /= np.std(discounted_epr)
epdlogp *= discounted_epr # modulate the gradient with advantage (PG magic happens right here.)
grad = policy_backward(eph, epdlogp, epx)
for k in model: grad_buffer[k] += grad[k] # accumulate grad over batch
# perform rmsprop parameter update every batch_size episodes
if episode_number % batch_size == 0:
for k,v in model.items():
g = grad_buffer[k] # gradient
rmsprop_cache[k] = decay_rate * rmsprop_cache[k] + (1 - decay_rate) * g**2
model[k] += learning_rate * g / (np.sqrt(rmsprop_cache[k]) + 1e-5)
grad_buffer[k] = np.zeros_like(v) # reset batch gradient buffer
# boring book-keeping
running_reward = reward_sum if running_reward is None else running_reward * 0.99 + reward_sum * 0.01
print(('%d , %d , %f ') % (episode_number-1,reward_sum,running_reward))
if episode_number % 20 == 0: pickle.dump(model, open('save.p', 'wb'))
reward_sum = 0
observation = env.reset() # reset env
prev_x = None