Timeline for What exactly is the advantage of double DQN over DQN?
Current License: CC BY-SA 4.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 25, 2021 at 9:28 | comment | added | David | @mohit yes, this is because of the $\epsilon$-greedy exploration. | |
| Mar 25, 2021 at 7:47 | comment | added | Mohit | Interestingly DDQN is not going to 0%! | |
| Jul 31, 2020 at 13:26 | history | edited | David | CC BY-SA 4.0 |
added 18 characters in body
|
| Jul 31, 2020 at 11:40 | vote | accept | Chukwudi | ||
| Jul 31, 2020 at 11:04 | comment | added | David | Yes, exactly. However, your online/main value network is just randomly initialized... Depending on the method, they usually just initialize it to very small values, so, as I say, it is unlikely to have a strong effect. | |
| Jul 31, 2020 at 9:40 | comment | added | Chukwudi | But, at the beginning of training, isn't our target network just a copy of our main value network? What happens if those weights overestimate the values? | |
| Jul 31, 2020 at 9:30 | comment | added | David | Well, the target network is randomly initialized, so that is unlikely to be the case. | |
| Jul 30, 2020 at 22:10 | comment | added | Chukwudi | Suppose our target network is the one overestimating our Q values. We use our main value network for selecting the action and use our target for calculating Q values. Wouldn't this value still be overestimated? | |
| Jul 30, 2020 at 20:34 | history | edited | David | CC BY-SA 4.0 |
added 1962 characters in body
|
| Jul 30, 2020 at 20:08 | history | answered | David | CC BY-SA 4.0 |