Here is a time-series graph I made of some data I collected from automated, neural network runs of Super Mario World.  Plot.ly makes this super easy by allowing you to quickly select the axes and add new variables.  Here, I will be discussing some of Stephen Few’s best practices for time-series analysis and how they apply to this graph:

  • Aggregating to Various Time Intervals

I’ve chosen to display this data with the time interval being each “generation” of the neural net.  Increasing the time resolution to each species and genome creates a huge amount of noise in the data, in which trends get lost.  Going by each generation makes it much, much easier to see progression.

  • Viewing Time Periods in Context

Including the data from the start to the finish of each level allows us to see the entire progression of each generation of the neural network.  It’s easy to narrow down the time scale to make it look like there has been no progress made.  Plot.ly makes this very easy zoom into smaller time periods, but also allows easy access to the full context of the graph.

  • Optimizing a Graph’s Aspect Ratio

The problem here might be how I have my website set up (I’m working on a redesign), but this graph is very cramped.  Viewing it on Plot.ly’s site its much more comfortable.  Yoshi’s Island 4 was completed relatively quickly, but it’s hard to see on my site due to its location.

  • Stacking Line Graphs to Compare Multiple Variables

I’ve combined several levels of data to see how progression and the number of generations to completion vary.  Stacking the line graphs makes it much easier to read than having a set of individual graphs, though organizing it can create other readability troubles.

  • Expressing Time as 0-100% to Compare Asynchronous Process

I did not utilize this in my graph but I think it would be interesting to see.  Doing so could possibly uncover some level design patterns.  The Super Mario games are known for being designed in a way that introduces specific level mechanics early on, generally in a safe way, then ramps up the implementation of it.  I suspect that this would be evident when viewing the levels in a manner of percentage of completion.