Differentiable ABMs

ABMs are discrete

0

1

Individual updates may not differentiable

but aggregate dynamics may be!

Differentiable ABMs

\theta
\nabla_\theta \mathbb E[f(\theta)]

Enables gradient-assisted calibration such as Variational Inference

Differentiable ABMs

\nabla_\theta \mathbb E[f(\theta)]

How do we obtain this gradient?

Automatic Differentiation

Challenges

  • Discreteness of the ABM microstructure
  • Stochasticity
  • Memory requirements
  • Autoregressiveness (vanishing gradients)

Ongoing research

ICAIF

ICML

AAMAS

ICLR

JOSS

arenas slides

By nickbishop

arenas slides

  • 63