Double DQN: The Decoupled Target
Intermediate
The problem, from first principles
Q-learning's max over noisy estimates is biased upward, so it overestimates values. Double DQN (van Hasselt et al., 2016) decouples the choice of action from its evaluation: the online network selects the action and the target network scores it.
From theory to code
Implement double_q_target(r, gamma, done, q_next_online, q_next_target) using that decoupled rule.
Constraints
- Ties choose the lowest action index.
Hints
Hint 1
Find a = argmax(q_next_online), then use q_next_target[a] in the Bellman target.
Keywords
research-papersreinforcement-learningdeep-q-learning
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.