I am having a hard time understanding why my loss is constantly a zero when using DQN. I'm trying to use the gym environment to play the game CartPole-V0.

In the code below, r_batch indicates rewards sampled from the replay buffer, and similarly s_batch, ns_batch, and dones_batch indicate the sampled state, next states, and if the "game" was done. Each time, I sample 4 samples from the buffer.

# This is yi
target_q = r_batch + self.gamma * np.amax(self.get_target_value(ns_batch), axis=1) * (1 - done_batch) 
# this is y
target_f = self.model.predict(s_batch)
# this is the gradient descent computation
losses = self.model.train_on_batch(s_batch, target_f)

The values of target_q are:

[ 0.42824322  0.01458293  1.         -0.29854858]

The values of target_f are:

[[0.11004215 1.0435755 ]
 [0.20311067 2.1085744 ]
 [0.413234   4.376865  ]
 [0.24785805 2.6716242 ]]

I'm missing something here, and I don't know what..


OK, a bit more digging led me to this:

The TD target or "target value" gets its name because by updating a Q table or training a NN with it as a ground truth, the estimator will output values in future closer to the supplied value. The estimator "gets closer to the target".

So, I have to update the prediction network with the "ground truth".

    for i, val in enumerate(actions):
        predict_q[i][val] = target_q[i]
| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.