Agent-based models (ABMs) are powerful and versatile tools, used in fields as varied as epidemiology and traffic engineering. With the ability to map individual agents to entities in the real world, ABMs can articulate and even predict patterns of behavior with surprising accuracy.

But a longstanding criticism of ABMs is that the mathematical foundation underlying these models is not very rigid, said Dr. Tomek Loboda, a postdoctoral fellow in the University of Pittsburgh’s School of Computing and Information (SCI).

“One problem even proponents of ABMs will admit is that the way you make those models is you tweak the parameters to your liking,” he said. While nothing is wrong with that approach, Loboda said, “There aren’t any axioms or bounds that guide your choices.”

An ongoing research project led by Dr. Paul R. Cohen, Founding Dean and Professor at the University of Pittsburgh School of Computing and Information, along with Dr. Mark Roberts in the Graduate School of Public Health and Dr. Greg Cooper in the Department of Biomedical Informatics, aims to change that.

Titled “Curating Probabilistic Relational Agent-based Models,” the project in 2018 received an award of nearly $1 million from the Defense Advanced Research Projects Agency (DARPA).

These probabilistic relational agent-based models, or PRAMs, seek to put ABMs on a firmer mathematical footing, Loboda explained, while preserving the considerable modeling strengths of ABMs.

Significantly, apart from making the models more sound, PRAMs’ probabilistic foundation facilitates integration of data, text, and human expertise in machine-assisted curation of ABMs, which is one the project’s goals.

Imagine, Loboda said, making a model of a flu epidemic in the City of Pittsburgh. That’s a very specific scenario, one that an ABM could handle well. But now imagine you want to create a flu-epidemic model for a city the same size as Pittsburgh that can then be applied to comparable metropolitan areas. How would the model need to adapt to fit Cincinnati or St. Louis? And is it conceivable to use these models to create a conceptually higher-level model that can show us what a flu epidemic in any city might look like?

PRAM paves the way for these possibilities, Loboda said, by allowing a machine to reorganize and reorder the modeling primitives, the models’ building blocks, to go from that lowest level—the Pittsburgh-based model—to higher levels.

That potential is in large part what drew DARPA to PRAM. The project is funded under DARPA’s Automating Scientific Knowledge Extraction (ASKE) program, which seeks to develop technologies capable of automating the manual processes of scientific knowledge discovery, curation, and application.

One possible application of PRAM that DARPA is interested in, Loboda said, involves helping scholars keep up with fast-growing bodies of literatures in their fields. Compiling summaries of relevant articles could help scholars to better stay abreast of new developments, effectively automating a time-consuming manual task.

In addition to providing the chance to incorporate machine learning, PRAM’s probabilistic foundation allows the modeling framework to incorporate several models simultaneously.

“PRAM is pretty flexible in allowing different modeling frameworks,” said Loboda. Rather than helping to build large, complex models that incorporate every facet of a situation, PRAM provides the ability to elegantly integrate several smaller models.

Think about dispatching an ambulance within a city, Loboda said. Such a scenario may be most clearly modeled by imagining several complex systems laid over one another: the ambulance and its relation to the hospital system, laid on top of the city’s traffic patterns, laid on top of a complicated road system.

“Systems are often largly separate,” Loboda said. “They’re layers on top of each other but they still interact.”

While the probabilistic aspect of PRAMs is key, Loboda also emphasized the models’ relational side as a central feature in its expressive power.

Returning to the example of tracking a flu epidemic, he suggested that a PRAM aimed at tracking flu transmission could use agents to track the numbers of people susceptible to, infected with, and recovered from the disease (or the SIR model, well known to epidemiologists), but also do so in connection to sites: a building on campus, for example, or a public park. A researcher might compare rates of infection in one dormitory versus another, or in different parts of town, while still being able to aggregate over those sites if necessary.

“But sites in Pram also stretch that definition to include, for instance, political positions,” Loboda added. A site isn’t merely a physical place but a kind of fixed position, in relationship to which PRAM can situate individuals. He recently spoke with a political scientist, for example, about using PRAM to explore how exposure to ads or certain kinds of information can affect political opinions.

It’s another layer of information and complexity that can help provide models even more potent and robust than the ordinary ABM. For example, answering questions about populations and subpopulations, or the so called lifted inference, is faster. And those are the sort of questions that inform policy the most.

The DARPA project is set to conclude in April 2020, following which additional projects are expected.

“As a research project, it’s good for PRAM to have potential applications, but we don’t know where it’s going to be tomorrow,” Loboda said. A large part of PRAM’s value at this stage is scientific, helping to shed light on a number of technical questions about PRAM and the nature of modeling. But as the popular quote goes, nothing is more practical than a good theory and PRAM promises to be a catalyst to the modeling process.

In that respect, Loboda sees Pitt’s SCI and the Modeling and Managing Complicated Systems (momacs) Institute as “the perfect habitat” for PRAM.

“PRAM is more than just technological machinery,” he said. “It’s a way of thinking that has the unification of modeling practices at its heart.”

That’s key, Loboda said, because a large part of what makes PRAMs’ potential so exciting is the prospect of aiding scholars and policymakers who may be less comfortable with modeling technologies.

“If we cannot understand, we cannot unify,” he said. “If we cannot unify, we cannot simplify, and if we can’t simplify we can’t enable non-modelers to model simply.”