Short Talk on Embodied Agent-based Modeling at CCS 2023

Last month I presented an idea that I have been working on for some time at the Conference on Complex Systems. Due to a last minute covid scare, I was unable to attend the conference in person. However, I was able to give a short talk which was titled Embodied Agent-based Modeling: Integrating Large-Language Models within Traditional ABMs. This is still an ongoing project and I will be releasing more information, including code, in the following weeks. For the moment, let me give you some background and whet your appetite for what is to come. You can find the slides for the talk here

Background

In most agent-based models, agents operate according to behavioral rules that dictate their responses to environmental changes or new information. These agents, that we term computational agents, are abstract, simplified versions of real-world agents. They are encoded within models, their behaviors and interactions guided by parametrized, stylized mathematical functions, with a focus on key aspects of behavior relevant to the model’s objective.

We extend traditional agent-based modeling and introduce the notion of an embodied agent. An embodied agent is an enhanced representation of a real-world agent within a simulated environment, aimed at capturing a broader scope of behaviors than those typically represented by a computational agent. While computational agents express behaviors that are computationally tractable, embodied agents aim to find a balance between algorithmic simplicity and the complex behaviors exhibited by real-world agents. These agents can process diverse forms of real-world information. This in turn is used to drive their interaction with other agents.

30000 feet view

In this work, I showed how embodied agents can be implemented within traditional ABMs. This is achieved by leveraging large-language models (LLMs). While LLMs have been used in a variety of contexts such as game-building and creating believable proxies of human behavior, their usage as a key ingredient within a tractable model of human behavior remains unexplored.

A schematic for integrating ABMs and LLMs is shown in the figure above. Using LLMs, we can continuously integrate unstructured data such as news feeds into the ABM. This is achieved through the LLM information layer. Each agent node $i$ within the ABM is then endowed with a local LLM which uses the information layer to drive the agent’s behavior. The interaction terms $r_{i,j}$ are then determined by the specificities of the particular ABM.