How do I update the Q values of a Deep Q Network when exploring?

Question

I am trying to implement a Deep Q Network to play Asteroids. Unfortunately, I am not sure how to calculate the Q value exactly, if I am exploring. For example, the agent is exploring for 1 second (otherwise makes no sense; I cannot let it just explore one step). Unfortunately, it makes a mistake at 0.99s, and the reward collapses.

At the moment, I am using the following formula to evaluate or update the Q value:

$$Q_{new,t} = reward + \gamma Q_{max,t+1}$$

But how do I know the max Q value of the next step? I could consider the best Q value the network says, but this is not necessarily true.

You can see the current implementation at the following URL: https://github.com/SuchtyTV/RLearningBird/blob/master/src/main/java/rlgame/Brain.java.

Answer 1

For tabular Q-learning, the q-values for state s and action a are updated according to

$$ Q(s, a) \gets Q(s, a) + \alpha [(r + max_{a'} Q(s', a')) - Q(s,a)] $$

where $\alpha$ is the learning rate and $(r + max_{a'} Q(s', a')) - Q(s,a)$ is the difference between the current estimate of the q-value, $Q(s,a)$, and the target, $r + max_{a'} Q(s', a')$.

The target q-value is based on the greedy policy, not the exploratory policy. Q-learning is theoretically guaranteed to converge to the optimal policy for any behavior policy (like $\epsilon$-greedy) that is guaranteed to visit every state and action pair an infinite number of times. See Section 6.5 of the Sutton and Barto book for more details.

In contrast to Q-learning, the target q-value for SARSA is $r + Q(s', a')$, where $a'$ is chosen from an exploratory behavior policy like $\epsilon$-greedy. For SARSA the learned q-values are dependent on the behavior policy and therefore not guaranteed to converge to the optimal policy. A behavior policy that intentionally acted randomly for multiple consecutive actions, as in your example Asteroids exploratory policy, would likely lead to learning different q-values than would be learned for an $\epsilon$-greedy behavior policy.

Unfortunately Q-learning's theoretical guarantees of convergence to an optimal policy go out the window when nonlinear function approximation is introduced, as is the case for deep neural networks. Nevertheless, in the Deep Q-Networks paper, the q-value function is updated using a target value based on the maximum q-value for the next state. Specifically, if $Q(s, a, w)$ is a q-value function parameterized by weights $w$, then the weights are updated by

$$ w \gets w + \alpha [(r + max_{a'} Q(s', a', w^-)) - Q(s, a, w)] \nabla_w Q(s,a,w) $$

where $w^-$ are the parameters of the target network used to stabilize training. (See the paper for more details). This update rule is chosen to minimizes the loss function

$$ L(w) = E[(r + max_{a'} Q(s', a', w^-)) - Q(s, a, w)]^2 $$

For your own implementation, it may be helpful to see a code example of the Deep Q-Networks parameter updates. A tensorflow implementation is available in the function build_train in the OpenAI Baselines DeepQ code.