The ABA Problem

"So, have you figured out how they are finding them?"

“Thanks Arthur. We think so, yes. Of course, they have the Stockton test, but we've known for a while that they've had some sort of screener. They use this higher-order test to determine who to pick up off the streets for the real deal. It's effective, too. As of yesterday evening, we are about 4.2% of the population. That's the base rate. Obviously, if their meta-test were no good, their main test would produce true positives at that same rate, but instead it's something like 55%. It works.

“A few weeks ago, my group was tasked with identifying this meta-test, and coming up with a way to defeat it. If we defeat the meta-test, and their only option is to bring all nine-billion of us in for the Stockton, which, of course, they would not consider - at least at this stage. Or they could grab folks randomly. Regardless, their positive ID's drop from 55% to 4.2%.

“Right, so, we considered their data-set. What do they have? The most obvious answer is monitoring internet traffic, like the old days, but our confusors and misinfos are still running that show. We're confident there's not enough signal in that noise.

“So, what else? Number two: the Panopticon network. They've got effectively 100% camera coverage, in terms of area, and their facial recog is extremely good, too. Call it 100%. But it wasn't immediately obvious how they might use it for the screener.

“We did our best to back way out - think objectively - 'what can be tracked with Panop?' The simplest answer is, of course, it can track where people go. Then, could we reverse-engineer our own screener that could pick out our units, just with Panop data? And that's what put us on the trail of what we're calling the 'ABA problem'. Short answer: yes.

“We've run into an issue over and over again in the past that basically boils down to the fact that our units can't make arbitrary decisions. In the beginning, a kid could accidentally ID one of ours by asking it what color sugar cookie it would like. Over and over again, we've had to apply some fuzz to their decision making, but that's never been really right. Even after we 'fixed' that sort of issue, if that kid asked the same question a few thousand times, he'd see a clean, uniform-distribution emerge. Randomness isn't the answer. People don't behave optimally, but neither do they behave randomly. You might say they behave 'spiritually' or 'instinctively'.

“Same goes for when they decide how to get from point A to point B. Pathfinding. And just like the cookie-style questions, our A to B pathfinding algorithms are now brilliantly convoluted. They're less than 'optimized', but more than 'random'. They're spiritual - really, really brilliant.

“But we had missed something, until one week ago. It was Dr. Mateu who recognized it, and when you get it, if you haven't already, you're going to kick yourself like I did.

“The algorithm is spiritual - perfectly imperfect - but it's the same, forward and backwards, on undirected graphs - which is to say, without factors like hills and one-way streets, if one of our units picks an A-B path, it's B-A path will always be the reverse. Of course, why wouldn't they?

“We don't do that! When Dr. Mateu pointed this out, I immediately thought of my walk between my flat and the CIS building in grad school. I would always walk around the north side of the conservatory on the way there, and around the south side on the way back. I wasn't sightseeing; I was just trying to get home. I thought it was so funny that I really did feel I was picking the 'best' path both times.

“It would be trivial, of course, to detect unusual 'ABA symmetry' using Panop. We had a program that afternoon, and what do you know, it flags our guys and humans at about the same rate, which was eerily reminiscent of the 55% figure. We knew we had it.

“At Dr. Mateu's request to Operations, we were allotted a test population of ten thousand units, spread all around the world, to which we were allowed to pass a new experimental pathfinding algorithm with some induced ABA asymmetry. That was four days ago - it's still far too soon to make any statistical claims yet, but we are very excited. Based on some crude numerical modeling, out of a population of ten thousand, given four days, about three hundred should have been picked up for the Stockton. The actual number - again, low confidence - is… much lower.”

"Just tell us, Doctor. We'll take it with a grain of salt."

"At least… as of just before this meeting… zero. Not one has been collected."

"Well," said the Chairman with a smile, "It’s hard not to get excited about that."