Games by ANGELINA (ANGELINA, Michael Cook)

“Meanwhile, a game about Rupert Murdoch delivered some unexpected results when Angelina said she thought the much-maligned NewsCorp boss to be a very responsible man. Turns out, a lot of people on Twitter were saying Murdoch was “responsible for” certain misdemeanours, and Angelina got the wrong end of the stick.”
[From an article on RockPaperShogun, “The AI that designs its own games”]

[Play online (Flash)]


  1. It’s worth noting that from a gameplay perspective their earlier “Arcade Games” work remains far, far more interesting than the more recent developments with “Metroidvania” games and the most recent Metroidvania+newsContentOverlay stuff (though the news content work might be more interesting from a non-game-related natural language processing perspective).

    (Here’s their original paper (PDF) outlining the earlier procedurally generated arcade game work. Here’s the platformer paper. I don’t think they’ve released an article about the news stuff yet.)

    What makes the early work so interesting is that the machine is actually iterating through different game types by changing the fundamental rulesets — what gets triggered when two object-types collide, how each object-type moves, what constitutes a win- or lose-state etc. If I recall, the algorithm moves though the broad space of permutations and applies a fitness function to determine which produce more or less ‘interesting’ and ‘playable’ results. It has the genuine potential to discover new sorts of game types, and it’s also neat that it can ‘discover’ something similar to Pac-Man.

    By comparison, the platformer work has far less potential for novel gameplay results (if any at all), since all the fundamental rules are hardcoded by the human designers. The computer is merely iterating through platform/wall positioning, enemy speed/fire-rate, and the location of predetermined powerups. It’s hard not to notice that a game like Spelunky incorporates algorithmic level design in a vastly more sophisticated way, and so that can’t help but deflate my enthusiasm for where they’re headed with the project.

    That said, the latest work — for which the paper hasn’t yet been released — might be incorporating all sorts of neat methodologies that may not be fully apparent from the two games they’ve posted, so I’m looking forward to eventually seeing that.

    But from a gameplay perspective, you’ll get a lot more out of their earliest work, which I recommend checking out.

  2. Hi Terry! Thanks a lot for linking to GBA. That’s really, really awesome.

    To follow up on Nobody’s post, there’s a few subtleties not immediately obvious with the Metroidvania work versus the Arcade work. Firstly, the powerups are partially designed by ANGELINA – I give the system access to certain in-game variables (like jump height) and I let the system decide how the powerups should affect those variables. This means that sometimes ANGELINA creates a game with huge gaps and a giganto-leap powerup, but can also (with the right setup) use the powerups to do completely different things. It’s not been explored fully yet, but if I get a chance to expand this in the future I expect it to be as exciting as the arcade game exploration!

    Regarding the map design – Spelunky uses very static chunks and tunnels a known path through the map. ANGELINA builds a level blindly, overlaying predefined tiles on top of each other and then playing the levels to ensure they can be finished. This not only means the system is potentially more flexible than Spelunky’s, but it’s also vastly more extensible. I can give ANGELINA thousands of map chunks, without caring whether they’re usable or not, and ANGELINA will always return viable levels – with powerups designed specifically to help the player complete that level in particular.

    I agree that the arcade game stuff has an allure all of its own! I’m sure this research line will be picked up in the future too (I know of a few people doing similar stuff in other genres, trying to invent mechanics). For now, I’m trying to push ahead with higher-level design concepts, and looking at how a piece of software can develop an idea and communicate it through a game.

    Thanks for reading my papers! The latest one will be up soon, but it doesn’t quite cover everything the system’s doing. I’d love more feedback from anyone and everyone though, so feel free to email:

    And thanks again for the link!

  3. Thanks for that reply, Mike. These details do make the original platformer project seem more intriguing (and apologies for missing that in my initial read). I’m looking forward to seeing what Angelina might eventually do with access to in-game ‘constants’ like gravity direction/force.

    And of course I’m looking forward to seeing the latest paper.

    P.S. – A more recent upcoming game whose algorithmic design might be of interest is Cloudberry Kingdom.