How to Perform Gradient Descent for DQN Loss Function

2 views (last 30 days)

Sherry X on 10 Mar 2020

Edited: Sherry X on 10 Mar 2020

I'm writing the DQN from scratch, and I'm confused of the procedure of updating the evaluateNet from the gradient descent.

The standard DQN algorithm is to define two networks:

. Train

with minibatch, and update the

with gradient descent step on

I define

. When update the

, I first make the

, and then only update

, which guarantee the

. Then I update the

. If I choose the feedforward train method as '

', does [1] update the evalNet correctly via gradient descent?

Find the treasures in MATLAB Central and discover how the community can help you!