Your reasoning is correct, I think this is just a typo. Indeed, for policy gradient algorithms, we want to perform gradient ascent to find the policy that maximizes the total expected discounted reward. A quick check is always to look at the consistency of the update as you did. If the TD-error is positive as well as the derivative of the log-policy, we have a positive ∆θ, which is what we expect.
- Contact
- EPFL CH-1015 Lausanne
- +41 21 693 11 11
Follow the pulses of EPFL on social networks
© 2023 EPFL, all rights reserved