making complexity simple
differentiable learning over millions of autonomous agents

Large Population Models (LPMs) are grounded in state-of-the-art AI research, a summary of which can be found here.

AgentTorch LPMs have four key features:

• Scalability: AgentTorch models can simulate country-size populations in seconds on commodity hardware.
• Differentiability: AgentTorch models can differentiate through simulations with stochastic dynamics and conditional interventions, enabling gradient-based learning.
• Composition: AgentTorch models can compose with deep neural networks (eg: LLMs), mechanistic simulators (eg: mitsuba) or other LPMs. This helps describe agent behavior, calibrate simulation parameters and specify expressive interaction rules.
• Generalization: AgentTorch helps simulate diverse ecosystems - humans in geospatial worlds, cells in anatomical worlds, autonomous avatars in digital worlds.

AgentTorch is building the future of decision engines - inside the body, around us and beyond!

Install the framework using pip, like so:

> pip install git+https://github.com/agenttorch/agenttorch

Some models require extra dependencies that have to be installed separately. For more information regarding this, as well as the hardware the project has been run on, please see docs/install.md.

The following section depicts the usage of existing models and population data to run simulations on your machine. It also acts as a showcase of the Agent Torch API.

A Jupyter Notebook containing the below examples can be found here.

# re-use existing models and population data easily from agent_torch.models import disease from agent_torch.populations import new_zealand # use the executor to plug-n-play from agent_torch.execute import Executor simulation = Executor(disease, new_zealand) simulation.execute()

# agent_"torch" works seamlessly with the pytorch API from torch.optim import SGD # create the simulation # ... # create an optimizer for the learnable parameters # in the simulation optimizer = SGD(simulation.parameters()) # learn from each "episode" and run the next one # with optimized parameters for i in range(episodes): optimizer.zero_grad() simulation.execute() optimizer.step() simulation.reset()

from agent_torch.llm.qa import SimulationAnalysisAgent, load_state_trace # create the simulation # ... state_trace = load_state_trace(simulation) analyzer = SimulationAnalysisAgent(simulation, state_trace) # ask questions regarding the simulation analyzer.query("How are stimulus payments affecting disease?") analyzer.query("Which age group has the lowest median income, and how much is it?")

A detailed explanation of the architecture of the Agent Torch framework can be found here.

A tutorial on how to create a simple predator-prey model can be found in the tutorials/ folder.

Thank you for your interest in contributing! You can contribute by reporting and fixing bugs in the framework or models, working on new features for the framework, creating new models, or by writing documentation for the project.

Take a look at the contributing guide for instructions on how to setup your environment, make changes to the codebase, and contribute them back to the project.

AgentTorch models are being deployed across the globe.