Algorithmic prediction

The nice thing about the way we’ve predicted for centuries is that it lets us skip ahead. That, after all, is the point of prediction. But our new form of prediction won’t take any shortcuts. And that makes it weird … but also, perhaps, a more accurate representation of how the future works.

With the old way we predict by figuring out a law of nature—a regularity that can be captured in a mathematical formula. Gravity is like that. If you drop a coin from the Tower of Pisa, we can predict when it will hit the ground because Newton figured out the formula for the attraction between two masses. Of course, we have to figure in the resistance of the atmosphere, but there’s a law for that as well. Plug in the numbers and you get your prediction. And it will be right.

Of course, if you want your prediction to be really really accurate, you’ll also have to figure in the current temperature because that will affect the density of the air. Also, better check the wind. Also, make sure no pigeons are flying by that could knock the coin off its course. Also, check for any seismic changes that might alter the angle of the tower, affecting the distance the coin falls.

None of this should cause us to devalue the awesome ability that formula-based prediction has bestowed on us. And as our engineering has gotten better, we’ve been able to use our predictive powers to land humans on the moon that circles us, to wheel robots on Mars and to enable spacecraft to escape our solar system.

At the same time, we should recognize that formula-based prediction has strong practical limits. We can predict where Voyager is at any moment, at least within some degree of accuracy, but we have no idea when a tiny fragment of space rock might slam into it. We don’t even have a lot of confidence about when an asteroid might head earthward so swiftly that we can’t stop it.

Game of Life

Algorithmic prediction works differently. A model for it is the Game of Life invented by John Horton Conway in 1970. It sets up a few arbitrary rules for when a square on a grid will turn black or white based on the state of its eight neighbors. If fewer than two neighboring squares are white, the square turns (or stays) white. If there are exactly three, it turns black. And so forth. You start out with an arbitrary set of squares filled in and apply the rules. You can do this without a computer, but it’s boring and painstaking—just the sort of task we invented computers for. (Cellular automata, intensively researched by Stephen Wolfram, play with the rules themselves.)

If you use a computer and put each design through dozens or hundreds of iterations, most of them result in random, pattern-less variations. But some show surprising regularities. One design seems to flap its “wings” as it moves up the grid. Another shoots out “bullets.” But, except perhaps for some of the simplest of them, the behavior of the designs is quite literally unpredictable. Yet all we’re talking about are literally black-or-white decisions based on a handful of incredibly simple rules.

Even with something as simple as the Game of Life—two states, governed by a small handful of rules—we usually can’t predict what the outcome will be. We have to step through it, iteratively applying the rules. If you want to know what a particular pattern will be after a hundred iterations, you’ll have to do the hundred iterations. You can use a computer to do this and it will spit out the answer, but it too has to step through the algorithm a hundred times.

Both forms of prediction work, of course. Formulas like Newton’s let us predict events on the assumption that nothing intervenes and that we don’t care about tiny inaccuracies. For example, the Earth isn’t the only object exerting a gravitational pull on a falling coin. So does the moon. And the sun. And everything else. On the other hand, algorithmic predictions work when we can get enough data in and when the rules are simple enough. But in real life they never are. For that you need the Game of Life.

If each type of prediction has its place and its limitations, we can still ask: Which form of prediction more closely models how things work? For me, formulaic prediction feels like a convenient shortcut, while algorithmic seems to be closer to the impossibly complex clockwork that actually drives change. And that makes the universe feel different.