# Questions on reinforcement learning topics would be better suited for AI Stack Exchange

I've notice that there are several unanswered questions on RL topics on this website. More specifically, about 1/3 (221/617) of the questions on RL on this website have not received an answer.

This also means that there isn't enough competent people on this website to answer such questions.

Apart from the fact that I think that RL is the central topic in AI, I think, all RL questions should belong to AI Stack Exchange website.

So, we need a way to bring these questions to https://ai.stackexchange.com/, where there are, at least, 2 competent people which could potentially answer such questions. Is there an automatic way of doing that?

What do you think about it?

• "better suited" is not especially relevant for decided whether something stays or is migrated; if it falls within our scope, we have no basis to move it unless the OP asks us to. i.e. we'd have to declare such posts off-topic before it would be acceptable to do anything the OP didn't ask us to (and I don't presently see a good reason to reduce our scope below what it was before AI got made). If it's on topic it's up to the poster to choose where they want to post their question. If there are participants at AI that can answer questions getting posted here, why not do so? – Glen_b Feb 13 '19 at 0:22
• @Glen_b I just think that websites should now start to focus more on specific topics (given the creation of new websites to address more specific problems). For example, if I am just interested in RL, I don't want to have multiple accounts and visiting multiple websites only because RL questions are on topic on multiple websites. This process of clear separation not only will help askers but also answerers. – user82135 Feb 13 '19 at 0:35
• If you're calling on us to reduce our scope, you should probably focus on that. Stats questions pop up on math.SE, stackoverflow, mathoverflow (and a couple of others) but I don't ask them to reduce their scope to suit me; as long as it's on topic for their site, I can't say much. There's naturally going to be overlaps. – Glen_b Feb 13 '19 at 0:38
• About a third of all questions on this site are unanswered. According to your logic, "there [aren't] enough competent people on this website" to answer questions about stats or machine learning, period. Did you really intend to offend the many people on our site who do regularly answer RL questions by insinuating that no more than one of them is "competent"? – whuber Feb 13 '19 at 0:50
• @whuber I did not intend to offend anyone, even though that statement is a fact, I didn't mean anything more than the fact itself. I actually thought (but I didn't write it) that the situations regarding the unanswered questions could also apply to other topics and websites. I just think that questions related to RL are better suited for ai.stackexchange.com. That's it. In fact, research in theoretical AI is mainly RL, nowadays. I'm not saying that there isn't an overlap, but it would have been better that most of those questions were asked on AI SE. – user82135 Feb 13 '19 at 0:56
• You could comment on OP's post and suggest that there is Artificial Intelligence that might be more suitable. However, as long as this community deems RL as on-topic, then no one can force all questions to be posted/migrated to there, unless this community decides to make it explicitly off-topic on here (just as Glen_b recommended). – Andrew T. Feb 13 '19 at 5:14
• I guess that what we are saying in here is that new stackexchange.com pages appear constantly and we have no control on this. If each time new page appears all the other pages needed to migrate all the questions that fall into scope of such page, we would be spending a lot of time on migrating questions between the pages. Scopes of different pages overlap. – Tim Feb 13 '19 at 9:59
• Does this question really deserve $-7$ downvotes (and, I guess, they will increase)? Is the quality of the question really so bad? This question is actually somehow similar to stats.meta.stackexchange.com/q/1117/82135, which received a lot of upvotes. – user82135 Feb 13 '19 at 16:16
• Downvotes on Meta have different flavour. They are to be construed as "I disagree with this proposal", not as "This is a lousy question". Now I have trouble remembering where I first read that, but it's a frequent remark around here. – Nick Cox Feb 13 '19 at 18:25
• e.g. see meta.stackexchange.com/a/195157/202463 or the question here: meta.stackexchange.com/questions/270/… – Glen_b Feb 13 '19 at 22:06

When you consider the lower proportion of unanswered questions on AI overall, there doesn't seem to be much evidence that RL questions have a relatively higher chance of being answered there rather than here compared to other questions. I obtained counts using the search terms: -[reinforcement-learning] is:question closed:no isanswered:yes, etc., and conducted the following analysis:

# input search data
x <- data.frame(      site=c("CV","CV","AI","AI"),
RL.tag=c(TRUE,FALSE,TRUE,FALSE),

# fit saturated logistic regression model

x  # print search data
# 1   CV   TRUE      341        242
# 2   CV  FALSE    71298      54360
# 3   AI   TRUE      251         83
# 4   AI  FALSE     1831        761

summary(sat.mod)   # print model summary
# Coefficients:
#                   Estimate Std. Error z value Pr(>|z|)
# (Intercept)        0.87798    0.04313  20.357   <2e-16 ***
# siteCV            -0.60674    0.04350 -13.947   <2e-16 ***
# RL.tagTRUE         0.22863    0.13376   1.709   0.0874 .
# siteCV:RL.tagTRUE -0.15692    0.15808  -0.993   0.3209

confint(sat.mod,level = 0.95)  # print confidence intervals around parameter estimates
#                         2.5 %     97.5 %
# (Intercept)        0.79392982  0.9630236
# siteCV            -0.69250867 -0.5219484
# RL.tagTRUE        -0.02926792  0.4957067
# siteCV:RL.tagTRUE -0.47007321  0.1500825


But I'd caution against reading too much into this kind of analysis—that doesn't control for the different types or varying quality of questions that get asked on the two sites. And tagging behaviour needs to be thought about; e.g., we've got 65 questions with but without . And who's to say that comparisons made now will be valid in a year's time, especially if there are only a few users answering RL questions on each site?

• Of course there are things to consider besides where a question's most likely to be answered, but I'll leave them for others, or for later. – Scortchi - Reinstate Monica Feb 13 '19 at 14:38

I think that the core of the issue is that https://ai.stackexchange.com/ and https://stats.stackexchange.com/ have way too much overlap. E.g., see:

If Stack Exchange wants create a website for each trendy name (AI, data science, etc.) without caring about how much they overlap with existing websites, then I'd suggest to implement "crossover questions" between sites.