Hey folks, thanks again for all the feedback. Just wanted to provide a quick update here about some developments in the last week, and about the overall trajectory of Radio.
@Geoff_Coupe’s report above gave me an idea, and we did some testing and found a bug 
Right now, Radio should only start based on content from your library. We are planning to remove this limitation soon, but for now TIDAL content that’s not in your library should not start Radio.
The bug causes Roon to start shuffling the entire library instead of playing similar content when the queue ends on a TIDAL track from outside your library. In this case you would see “Playing Music Similar To Your Library” as shown in some of the screenshots above. Obviously, these selections would not be related to what you were listening to, and could certainly can give the appearance of some very random selections. This issue should be resolved in our next release.
Now that’s a nice crisp bug, and one that we’ll be rolling out a fix for soon. That said, I doubt all the feedback in this thread is related to that one issue, particularly based on the other screenshots I’m seeing in this thread.
As I mentioned previously, ensuring that Radio always makes great picks is our goal, and since the previous algorithm was quite simplistic, we really consider this our first attempt. The new Radio algorithm is built to learn smarter behaviors over time, and also supports some very granular tweaking by our team in response to feedback we hear from you guys. Improvements will happen continuously over time, but I do want to be clear that low quality picks are different completely random selections caused by bugs.
Roon’s Radio functionality gets smarter all the time, but in general the selections are only going to be as good as Roon’s knowledge of the initial “seed” selection. If we don’t have genres for the initial content, the picks may feel less related. If the initial album isn’t identified, we may lack knowledge of the artist or performer, as well as their influences, related artists, and more. Unidentified albums may also lack credits and composer data that helps us correlate content.
It’s worth mentioning that if you’re only using your own genres instead of Roon’s, the selections may lack some of the “guardrails” that are built into the system. If your Genre grooming is incredibly detailed, you may end up with a great experience, but Radio is going to make better picks if it knows about 40 sub-genres of Jazz, as opposed to having everything just tagged as “Jazz”.
Additionally, Radio is designed to limit repetitiveness with regards to artists, performers, and albums. If your library is small, and you only have a handful of artists in a given genre, you may get broader selections as we branch out, rather than the same artists coming up over and over.
Radio may also be giving weight to factors you’re not aware of, or which don’t translate well in practice. In my early testing, I saw 80’s New Wave Punk content come up after starting with a current R&B track. It turns out the punk artist had been sampled on the R&B song and got a songwriter credit. So they were related, but the pick didn’t really work.
We made changes to the algorithm after noticing that, and we will continue to tweak the selections as reports come in. Again, the most important thing you guys can do is give me detailed reports including the information I mentioned above. Feel free to also provide us information about what picks from your library would’ve been more appropriate, or what types of connections you feel need to be emphasized more.
Hopefully this insight is helpful for everyone interested in helping us get this right. When reporting issues, please continue to provide details about what metadata Roon has for the seed, and for any bad selections you want to report back to us. Additionally, the more you can tell us about what would constitute a good pick for you, the closer we’ll be able to get.
Radio will continue to improve, so please keep the feedback coming. We really appreciate it.
