Timeline for How do we define greedy action for policy improvement for a given stochastic policy?
Current License: CC BY-SA 4.0
14 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 10 at 10:01 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Apr 12 at 10:01 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Dec 14, 2023 at 9:01 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Nov 14, 2023 at 8:05 | answer | added | cinch | timeline score: 0 | |
| Nov 8, 2023 at 11:00 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Jul 11, 2023 at 10:04 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Jun 11, 2023 at 11:23 | comment | added | Neil Slater | Other than that, if you are talking about policy improvement in Q-learning, or SARSA or other action-value based method, then the definition for greedy action/greedy policy does not change. If you are interested there is an extension to the policy improvment theorem that the resulting $\epsilon$-greedy policy (in e.g. SARSA) is a strict improvement over the previous one. | |
| Jun 11, 2023 at 10:45 | comment | added | Neil Slater | I think the question is missing some context. The greedy action is by definition the action that the agent predicts will result in the highest return. That's it, there's no edge cases or special rules beyond that in different circumstances. I'd like to know what you are trying to do with this information. For instance, it would not be appropriate to use the greedy action to train a policy gradient method - you should use the action taken by the policy even if it was "wrong". | |
| Jun 11, 2023 at 9:26 | comment | added | Luca Anzalone | Sure, have a look at my answer | |
| Jun 11, 2023 at 9:26 | answer | added | Luca Anzalone | timeline score: 1 | |
| Jun 10, 2023 at 19:05 | comment | added | DSPinfinity | Could you please write your guess mathematically? | |
| Jun 10, 2023 at 15:38 | comment | added | Luca Anzalone | Probably related: ai.stackexchange.com/questions/40149/…. In my opinion, the action that can be considered "greedy" for a stochastic policy, should be one of the distribution mode(s). | |
| Jun 10, 2023 at 15:22 | history | edited | Luca Anzalone | CC BY-SA 4.0 |
improved mathjax
|
| Jun 10, 2023 at 13:49 | history | asked | DSPinfinity | CC BY-SA 4.0 |