Functional ABM

Prototype Functional Agent Based Modelling API

Introduction

This is a prototype attempt to create a simple modelling API based on graph computation and functional patterns, where most ABM use OOP based patterns (probably rightly so!).

Though the resulting pattern is not strictly speaking functional (it ended altering the state of the model in place) it does provide a convenient API for quickly building models.

Usage

Several examples of a few basic models can be found in the examples folder

Agents are defined by an update function that defines how an agent updates it's state each time it is activated during the simulation.

Once decorated this function is wrapped as part of a recurrent computational graph connecting agents in a causal manner.

Update functions should follow this pattern:

def update(t, antecedents, state, descendants):
    # Perform updates
    return next_step

Where the arguments are:

• t: The current time/step of the simulation
• antecedents: Input states that precede this agent
• state: The state owned by (and updated by) this node
• descendants: Nodes downstream of this agent, that can also be altered by this node

the update function should alter the agents state (and any descendants were applicable) in place and then only return the next time/step the agent will be activated.

Decorating an update function wth the agent decorator creates a new type wrapping the update function and allowing states to be assigned as arguments to each instance of an agent using the update function.

Multiple agents types can be created using multiple update functions, and arbitrary structures used to hold the state of the model (numpy arrays, graphs etc.).