ML Newsnew | past | best | rss | submit about
Super Mario as a String: Platformer Level Generation via LSTMs (arxiv.org)
52 points by eusebio on May 27, 2016 | hide | past | pdf | 5 comments on HN

In plain words: Levels are written as strings of characters and fed to a memory-based text generator that learns to continue them, making new Mario stages. Unlike the usual approach that predicts each piece from the last few, its levels fell within the range of human-made ones.

Abstract · Super Mario as a String: Platformer Level Generation Via LSTMs

The procedural generation of video game levels has existed for at least 30 years, but only recently have machine learning approaches been used to generate levels without specifying the rules for generation. A number of these have looked at platformer levels as a sequence of characters and performed generation using Markov chains. In this paper we examine the use of Long Short-Term Memory recurrent neural networks (LSTMs) for the purpose of generating levels trained from a corpus of Super Mario Brothers levels. We analyze a number of different data representations and how the generated levels fit into the space of human authored Super Mario Brothers levels.

Adam Summerville, Michael Mateas
arXiv:1603.00930 · cs.NE, cs.LG · submitted Mar 2, 2016 · updated Mar 8, 2016
abstract · pdf · html

add comment on HN

jc4p on May 27, 2016 [–]
This is awesome! There's a lot of interesting things that can be done with just the level of sequence generation AI we currently have, transforming the real data into a format for it is the most difficult part, I think.

This similar project from a few months ago does Mario level generation and discusses the training data sequence creation in greater detail: https://medium.com/@ageitgey/machine-learning-is-fun-part-2-...

reply
panic on May 27, 2016 | parent [–]
Even cooler, you can actually play the AI-generated level from that article if you have Super Mario Maker: https://supermariomakerbookmark.nintendo.net/courses/4AC9-00...
reply
steven777400 on May 27, 2016 [–]
This would make for an interesting derivation of the Turing test: can a player determine if a given level was created by a professional game designer or AI?
reply
sleepychu on May 27, 2016 | parent [–]
If the corpus is human levels though then all of the interesting gimmicks are human though, no?
reply
rav on May 28, 2016 [–]
The authors define a level metric for the "linearity" of the level, defined as "how close the level can be fit to a line". This seems like a vague definition -- is it the optimal path taken by the player that is being fit to a line?
reply
About | RSS | RSS (all) | HN arXiv