Smart simulations
Mr. Lebiere’s team is developing the simulation for the SocialSim project. That is, they’re building a model that mimics human behavior, which they can then feed a scenario and examine the results (no humans required).
Imagine having a computer program that could analyze any situation and how the general public will respond — like a change in the Facebook algorithm. That’s the model.
To create it, the team will build a “cognitive architecture,” or reproduction of how humans think.
“[Cognitive architectures] are meant to represent, at some level, the computational processes that go on as we perceive the world, as we act on the world, as we make decisions and as we plan to solve a problem,” Mr. Lebiere said.
One successful example is CMU’s “Act-R” cognitive architecture, which was completed for release in 2014. It essentially breaks down the human mind into a series of operations.
Act-R has been used to create models in hundreds of scientific publications, and it’s been applied for use in cognitive tutors that mimic the behavior of a student — personalizing his or her curriculum and instructions based on any difficulties the pupil is experiencing.
The SocialSim simulation will rely on a huge amount of decision-making by millions of people — all without modeling each individual.
The challenge of SocialSim, in Mr. Lebiere’s opinion, is to develop a good understanding of human thought processes and then scale that to a societal level to see how large collections of individuals act as a collective. He offered an analogyto physics.
“You have a vat of gas, and you have molecules of gas,” he said. “The individual
The resulting model won’t be perfect, Mr. Lebiere said, but it will be useful. SocialSim should be useful in better understanding how U.S. adversaries are operating online, for one.
In a report, DARPA noted the U.S. government currently employs small teams of experts to speculate how information might be spreading online, but those teams work slowly, their accuracy is unknown, and ideas can only be scaled to represent a fraction of the real world.
Disaster relief campaigns could be more effective with a useful model of human behavior. The spread of rumors about U.S. military efforts in foreign nations can hinder the very help they’re providing, Mr. Lebierepointed out.
For example, a rumor that the bottled water the U.S. military brought to Haiti was contaminated with “mind control” substances could keep the survivors from drinking fresh water. If researchers understood how this false information spread, it might be possible to prevent it in the future.
In another scenario, a simulation of human behavior could allow the U.S. to solve conflicts without using violent force, he said.
And, of course, there’s always the possibility of understanding how fake news is spread — and stopping it.
In a commercial application, a company such as Facebook might use SocialSim to understand the real-world ramifications of a change in software.
“How is the algorithm change going to impact how the information spreads?” Mr. Lebiere asked. “The scale of human behavior is really playing out in this program.”