Wednesday, June 20, 2012

A Eureka Moment about Training, Education, Puzzles, and Games

I was thinking about a time when my department head came to my game design class unannounced to evaluate my teaching, and I wasn’t “lecturing” to the students.  They were working on game projects.  (This was not an introductory class.)  She seemed surprised that I wasn’t lecturing, but that may be because she typically taught introductory computer literacy style classes such as how to use Microsoft Office.  Classes that teach use of specific office software can be taught more or less by rote: if you want to make something bold you highlight it and press control-B or click the Bold button.  If you change margins you do thus and so.  And so forth.

These intro software classes don’t have to be taught entirely by rote but commonly they are, complete with what I call “monkey books”.  These books have students follow steps to accomplish something, but students tend to focus on getting through as rapidly as possible, and when they’re done they don’t know what they did and haven’t learned much.  Like the monkeys who, if they type long enough, type Shakespeare’s works . . .  You can learn from monkey books, but only if you want to learn and make the effort to learn.

Designing games is not and can never be taught by rote.  Teaching by rote is training, not education.  Education is about why you do things, why some things work and others don’t, about understanding what you’re doing.  Training is about exactly how you get a particular thing done.  I recognize that not everyone follows those definitions but I find it very useful to make this distinction, and other people with other purposes when defining education and training may make different distinctions.

Designing games is about education, not training.  Designing games is about critical thinking, and much of it is thinking, which is the antithesis of training.  You’re trained to do things automatically, without thinking.  (Reiner Knizia on twitter recently said, "To summarise my experience: Design is a way of thinking!")

Video game production at the outset can be taught by rote because people are learning how to use particular software, for example Maya or 3DS Max, or they’re learning how to program.  In the long run there is a process of education there, especially for programming, but in the short run for introductory classes a lot of it is simple straightforward “this is how you do it”.  There just isn’t much of that in game design.

But where the Eureka moment occurred was when I realized that an analogy can be made from this to games and puzzles.  A puzzle is something that has a solution, or perhaps several solutions, with the defining characteristic that once you figure it out the solution(s) always works.  So you can teach someone by rote how to beat the puzzle by teaching them the steps required.  It’s possible that those steps require certain skills such as hand-eye coordination levels that the person may not have attained, but once they attain those skill levels they can follow the solution and complete the puzzle every time, or as it is said in video games, “beat the game”.

A game does not have these kinds of solutions, and cannot be “beaten.”  To be good at the game requires something much more akin to education than training.  You have to understand why you’re doing what you’re doing, and when that isn’t the best thing to do, when something else is the best thing to do.  There is certainly problem-solving in games, but there aren’t solutions to the game as a whole that will always work.  Frequently this is the difference between having human opponents and having no opponent or a computer opponent, though computer opponents continue to become better over time.  Frequently this is the difference between, on th one hand, perfect information or uncertainty that can become predictable, typical in puzzles, and on the other hand uncertainty that cannot be predicted or accounted for by simple mathematical processes–the kind of uncertainty that comes from having several human opponents.

You can teach someone, by rote, how to win at Tic-Tac-Toe, or even Tetris, and you could for chess if anyone had completely solved the extremely complicated puzzle.  The checker program Chinook, as I understand it, plays by rote, playing what it knows to be the move most likely to lead to a win from whatever the current position is–no reasoning required.  You cannot teach someone how to win at Britannia or Dragon Rage, Diplomacy or even Risk, by rote, they have to understand how it all works and then think as they actually play.


Rafael Van Daele-Hunt said...

I'm not convinced that your puzzle/game distinction is valid. What is the difference between reasoning and rote within a formal system like a game? Hidden information and the intentions of other players can be modelled as part of the game state.

If checkers is a puzzle because there is a best move for a given board state, then Risk is a puzzle because there is a best move for a given board+cards+player psychology state.

Put another way: there is poker AI as well as chess AI.

Lewis Pulsipher said...

Barring a very simple game, there is *no* "best" move for a given boards+cards+player psychology state. IAW the mathematical theory of games, there is a range of possibly-best moves, with a mixed strategy for choosing one. If you have perfect information you may (in simple games) be able to calculate the best move, but with hidden information this becomes effectively impossible. Nor can you calculate player psychology of course, so Game Theory assumes an opponent who plays perfectly.

Checker and chess "AI" can use brute computing force to calculate a best move. A poker opponent using brute force calculation is unlikely to do well against outstanding poker players, despite the very simple game system in poker (much simpler than most game systems).

Rafael Van Daele-Hunt said...

I don't see that a game-theoretically optimal strategy of choosing (weighted-)randomly between a set of options is deeper or more interesting than a single best choice.

More importantly, I assert that player psychology *can* be calculated, or rather estimated statistically. Successful pokerbots do not play the game-theorical optimal strategy, precisely because although that would minimize potential losses, it fails to exploit flaws in opponents' play and hence doesn't maximize its overall winnings. Instead, they react to the play of their opponents.

In a sense, I'm agreeing with your claim: a game, as opposed to a puzzle, requires reacting to the other players. From a computational point of view, though, both are equally brute force.

For a human, games feel different from puzzles, for two reasons: first, because we really aren't capable of calculating a game's solution; secondly, because active opposition triggers competitive feelings that a passive puzzle doesn't.

I claim, however, that these are fuzzy distinctions: a sufficiently complex puzzle with the illusion of an opponent (as in a computer AI) doesn't feel different in kind from playing a "real game" against an opponent in another room.

Lewis Pulsipher said...

So you claim that computers have already passed the Turing Test? Yet video game players usually report that human opponents in video games are much better players, in general, than the computer game itself. One reason why I try not to use the term "AI" for a computer opponent is that it isn't AI, it's an approximation. Whether a true attempt at AI would do better may be true, but enough to approximate a human? I understand that computer experts feel the technological "singularity" is still quite a way into the future.

When I was young, I was a determinist, I thought that with a sufficiently powerful computer you could calculate everything that would happen in the universe. But as I got older I realized that this was infeasible regardless of technological level, and the Heisenberg Uncertainty Principle entirely put paid to the notion. Those who believe that all games are math usually believe that everything can be calculated. Those who (like me) believe that games are about people (puzzles are about math) think that full, infallible calculation is a chimera except in what amounts to a puzzle, such as chess or checkers, even though chess is too complex for humans to fully solve. (Marion Tinsley came close in checkers.)

Rafael Van Daele-Hunt said...

I don't think that computers have passed the Turing Test yet. I do think that they are close to it for many games. No Singularity is needed, since games (excluding table talk and diplomacy), are much more restricted in input and output channels than a true Turing Test.

Commercial video game opponents are *designed* to lose and to be predictable, because (as you note in other articles) most video gamers want an entertainment experience more than a challenge.

I realize that due to uncertainty, chaos, Goedel, etc. there are limits to computational prediction, but these limits are not necessarily reached in the constrained task of playing a game, even a game of hidden information.

If you would like to *define* as game as having a ruleset such that the best move for a given situation and set of players cannot be calculated exactly (including mixed strategies) within the timeframe of the universe, then I am willing to admit that such games exist. However, it would be easy to construct a perfect-information "puzzle" that fulfils this criterion (just scale Go up to a large enough board).

In short, I agree that imperfect information makes games for complex, for humans and computers alike, but disagree that there is a non-psychological difference in kind.

As an aside, my own tastes match yours -- I find the kind of games you describe as "games" much more satisfying than "puzzles".