Evolving an Understanding of Fun

An article by Julian Togelius was recently brought to my attention in which he discusses his experiments with a game generation EA that evaluates game fitness as a function of automated player training.

I think it's a great idea, if not a terribly obvious one. Julian's purpose in his quest is a noble one:

Our solution is to use learnability as a predictor of fun. A good game is one that is not winnable by a novice player, but which the player can learn to play better and better over time, and eventually win; it has a smooth learning curve.

I think if you ignore silly animalistic compulsions like collecting items, finding all the secrets, or acting out fantasies, as well as more legitimate human animal patterns like appreciating beauty in game audio or visuals, or social interaction through gaming (hopefully more like Super Smash Bros than World of Warcraft,) I believe this is an absolutely solid function for measuring game funness.

It's worth noting here that Togelius does not explicitly state that his technique is immediately applicable to video games, only heavily implied. (It should also be noted that by video games I am going to bravely exclude all turn-based games, relegating further discussion to games which challenge one to think fast and exercise their reflexes.)

What's really more important than whether or not the technique is applicable to just video games or just conventional games is whether or not it's applicable to both, because (and I think Togelius would agree) that says something very fundamental about games and what they mean. I say his techniques are indeed applicable to both, but I see some daunting technical hurdles before him if he wants to follow up on his insinuation that he can measure the fun of Pac-Man, and luckily those hurdles are not so nuanced as animalistic compulsions and enjoying artwork.

The problem is that even in a game which discloses 100% of all information in the game at all times, there is absolutely no reason to think this means that the player perceives 100% of information in the game at all times. In order to predict funness for a single person for a single video game, even with perfect knowledge of how well they learn to plan their moves, you need an accurate model of how well they learn basic things and utilize skills like ballistic extrapolation (or whatever movement dynamics are applicable to the game) for predicting collisions (for avoiding enemies and collecting items.) What is probably more nuanced about this is that a lot of the brain hardware is shared between seemingly disparate mental processes: e.g. having to focus more on collision prediction might distract players from path finding.

Those challenges notwithstanding, I believe the technique could provide a few valuable tools to both software and cardboard game publishers.

The first is that you could try to analyze different player types, by which I mean contrasting graduation pathways (simulated using alternate mutational pathways) on players' paths to greatness at a particular game. Pathways may diverge and recombine, or end up at entirely different play strategies for the same game. Never before have I encountered a game that really made room for players who might learn very differently from each other, and I think the application of Togelius' techniques might bring some formal validity and computer-aided design to those brave enough to write the first video game that dynamically tailors its tutorials based on player performance.

The second idea for application of Togelius' technique I'll begin to explain with an excerpt from his article:

A good game is one that is not winnable by a novice player, but which the player can learn to play better and better over time, and eventually win; it has a smooth learning curve. This can be seen as an interpretation of Raph Koster's "Theory of Fun for Game Design", or of Juergen Schmidhuber's curiosity principle.

Togelius' technique teaches an AI to play a particular game, and extrapolates game quality from the smoothness of the simulated player's ascent. I happen to think that there is more interesting and complex data in that learning curve than some scalar representation of its smoothness. I would wager that the technique is capable of simulating direct transfer of knowledge on game mechanics to the player, or if you prefer, instructing the player on how to play.

By examining the influence that each individual characteristic of player behavior has on the success of the simulated player, you could potentially identify different horizons of significance among those behaviors.

For instance, knowledge of some rules for a game may clump together "at the bottom" where the player's performance was the worst, demonstrating a necessity for an instructional component to a game. I don't believe this "kink" in the "smoothness" of the learning curve should be counted among the parts of the curve which represent the experience of playing the game after reading the instruction manual or playing through the tutorial level of a video game. (In fact, if Togelius' could work the generation of tutorial levels into his video games so that the automated player could train on a single challenge at a time before facing them simultaneously, he might iron out that kink in the learning curve data.)

Such a "clumping" analysis technique could be used to analyze the learning curve of existing games, providing game publishers with a new tool to determine what rules need to be part of a rule book. Sure there are certain expressions of game rules which are *required* to be in a rule book to express the full game, but there might be important emergent characteristics of game play appropriate for inclusion in rule books, or tutoral stages in video games.

Aside from the clump at the bottom, I think most games do (or should) have two other kinds of clumps as well, and I think these other two clump varieties may be even more important than anything else I or Togelius have discussed so far. The simpler of the two first: a clump at the top might materialize when even the best player cannot gaurantee perfection all the time. A little chaos element in a game goes a long way if you ask me. If you know you're going to win every time, the game ceases to be fun, and maybe that isn't simply a function of having run out of challenges to overcome, which brings us to our final clump variety..

When I was a child, when Megaman assimilated powers from boss enemies, it wasn't only rewarding because I had just defeated a unique and powerful opponent, or because I had indulged in the compulsion to collect powerups, but because it introduced a new horizon in game dynamics. Even without manufacturing artificial horizons in player strategy by holding out on the player, I believe a fundamentally fun part of games is not just overcoming unique challenges, but uncovering them! The most rewarding elements of game play aren't simply doing something new, but experiencing those Eureka! moments when you figure out something you could have been doing all along. You've discovered something innate, you haven't been given a hand out.

Interestingly I think this is my personal #1 goal when it comes to game design. I am most attracted to games which challenge the player to invent original solutons and offer highly flexible gameplay through a wealth of emergent game dynamics. I'm quite attracted to nerdy games which involve any kind of programming or sequencing causal chains to build complex and original devices that beg the question "Did the game developers even realize I could do this?" Many moons ago I was a big fan of the game Magic: The Gathering which had even named the canon of player strategies that involved complex original solutions made from the emergent properties between many components: the "combo."

I owe a debt of gratitude to Julian for writing the article which inspired me to fully appreciate the mathematical elegance of what it is I love about games, which in a word is: ingenuity.