Process
& Work
Curating Probabilistic Relational Agent-based Models
– Paul R. Cohen

Although agent-based modeling (ABM) is promising and widely used, agent-based curation is surprisingly primitive: With the notable exception of population synthesis methods, there are no algorithms to create large-scale ABMs semi-automatically, and ABM development frameworks make no contact with modern curation technologies such as ontologies, machine reading and machine learning. Worse, the semantics of ABMs are murky, in part because there are no curation tools to enforce semantics.

We will develop probabilistic models and algorithms for incremental, human-machine curation of ABMs and demonstrate curation in a disease outbreak problem and a long-term economic risk modeling problem.

The key technical challenges are:

  • Developing a probabilistic, relational target representation for agents that is accessible to and modifiable by humans, machine reading, and data mining algorithms;
  • Developing interaction protocols and interfaces whereby humans and machines can jointly negotiate over the structure and parameters of agents, while preserving the semantics of the target representation.

We will call upon diverse technologies to address these challenges, including population synthesis algorithms [4,20,21,24], probabilistic relational models and related learning methods [8,10], scant theoretical foundations for ABMs [2,7,12], machine reading for curation [1,13,23], domain-specific programming languages [e.g., 6], and ABM frameworks [11,16].

We will develop human-machine population synthesis methods that integrate machine reading and bring the full expressive power, inference methods and learning methods of probabilistic relational models to ABMs.

We are not proposing to build a new ABM framework. There are dozens, perhaps hundreds of frameworks for ABMs and their cousins [3,16]. The computational “guts” of ABM frameworks are more or less well worked out (particularly in our own Framework for Reconstructing Epidemiological Dynamics (FRED) framework [11]). What’s missing is support for curating ABMs.

If successful, the project would integrate data, text and human expertise in ABM curation. It would provide a clear semantics for agents and their behaviors. And it would accelerate ABM curation and maintenance.