Can AI predict setlists? Some thoughts…
Disclaimer: This is a translated version of a Dutch post I published in June 2017. Which is lightyears in the field of AI and machine learning I think? Some advances might have already been made.
Concerts are a wonderful experience of human togetherness. Whether you are at a classic piece in the theatre, at a starting punk band in the local pub or just like me, seeing Eddie Vedder at the start of his latest European tour, a concert is always a great night out.
Personally I am always very curious about the setlist of a concert. Will it be a regular set of greatest hits as you get on festivals or will the real fans get the night of their lives with b-sides, obscure tracks and new songs? Or is there anything in between? Would this be predictable with the technology that is becoming increasingly important in our lives: AI or Artificial Intelligence? How well can an algorithm with sufficient computing power and historical data predict a setlist for an upcoming concert?
Dynamics of a concert
Allow me to take a step back. I’m a great Pearl Jam fan. I’ve followed the band since their beginnings in 1992 and have been fortunate to see them live a number of times.
What attracts me to Pearl Jam is not just the music per se but also their live performances and the relationship with the fans. No concert is identical and they always go out to make every night special. Every night has its own dynamics and the band tries to respond to that with a unique setlist.
For example, you can watch two consecutive nights of Pearl Jam in Amsterdam and come home with 54 played songs of which 52 different ones. Because not only do they play many different songs every night, they play for a long time too. An average evening with Pearl Jam will take 2 to 2.5 hours easy.
Deadheads and robots
Sometimes different thoughts come together and create a new idea. This is one of those. I was reading some articles about AI and algorithms. For the sake of clarity, I am an absolute layman in that area. I do not know the tools and options sufficiently and I would not know how an algorithm actually works. For me, that’s a reason to learn about the possibilities and opportunities it offers. I read about beautiful future possibilities and especially in healthcare and safety, I see we are entering interesting times. But how does AI deal with creativity? With human inventiveness? That is the prevailing fear when I read about it.
Later that night I saw the trailer for the new (four-hour!) Documentary “Long Strange Trip” by The Grateful Dead.
This band from the sixties is known for their advanced music back then, their look on life and their concerts. Apart from the music style, there are many similarities between The Grateful Dead and Pearl Jam. Fans go all the way, music is a living organism to both the band and the fans and no show is the same. Best of all, both bands had the viewpoint: Record the audio of the concert and share it with others. There are official bootlegs from both Pearl Jam and The Grateful Dead, but they condone the very vivid subculture of trading and sharing your own home-recordings of the concerts.
Here’s The Idea
This all made me think. Pearl Jam’s setlists are hard to predict. Unlike other bands, they pick a different setlist for each show. They refuse to devaluate a concert tothe same list of hits and public favorites every night. They always dive into their rich archives to come up with surprises. A few songs always come back at regular times, but most of the times it’s a big surprise what’s coming.
Could an algorithm predict a setlist like that? If you get enough data from previous concerts, would it be possible to take a guess on the setlists for new concerts? And if you combine that with the sort of venue, whether it’s a festival, in what city or country is the concert and whether a new album has recently appeared?
Backstage magic versus server logica
I have no idea if it’s possible. To be honest, I hope an algorithm does not go beyond 40% of predicting the setlist of that evening. I still believe that the band can put together a more unique setlist than historical data will predict.
I think 6 band members backstage will be more creative than a mathematical model. It will be a mathematical model that has been fed with the creativity of the band over the years, but I really wish that it can’t accurately predict which songs will be played on any specific evening.
But what if it could? Suppose we have an algorithm that can predict the setlist for each evening of a tour? I honestly would not be happy and destroy the code sooner then later. A big part of a Pearl Jam concert is the surprise of the setlist. I would not want that to be ruined by an algorithm. But if it is, then we know that it’s possible. Mission accomplished? I don’t know yet…
How would such a mathematical model work?
I am an absolute layman in the field of algorithms, data processing and probability calculation. You can use the following approach to shoot, improve or change holes.
- Collect existing data from previous concerts. For example, at the Setlist.fm API or sites like http://livefootsteps.org/ or Pearljam.com itself.
- Collect data when albums are released.
- Process this data and create a model.
- Include how tours, new albums and setlists relate to each other. A fan has already made such a model as can be seen on the PJ20 DVD.
- Then we have to wait for a new concert series of Pearl Jam. There are rumors of a Latin tour in 2018…
- But maybe we can already test with other bands or other sets of historical data. Bands like The Black Crowes or The Grateful Dead? They’re done touring so maybe you can test the model?
- Will there be a new concert series? Based on the amount of concerts, places and dates, we will be guessing for the setlists.
- Include new material here when it’s available
- At each concert, just before the band hits the stage, we announce the setlist as the algorithm has designed it.
- During the concert we keep track of how we’re doing and where we made mistakes
- These are included in the forecast for the next concert
- Repeat the sequence.
Has this been done already?
I have not done an extensive study, but in a Twitter conversation with other fans and experts in the subject, ideas come up, among other things, about CreativeAI. But there have been previous attempts around Phish and (there they are again) The Grateful Dead. I want to do some more research in the coming period.
What’s next?
On Twitter we already had a furious discussion with experts and lovers and there are some people in my network who are intrigued by the idea. I hope to take a first step with some of them. I think I’d like to contact the owner of livefootsteps.org, as they have a fairly complete database. And after that? Let’s hear your ideas and see where it leads us!
