Technical reports

Scalable agent-based models for optimized policy design: applications to the economics of biodiversity and carbon

Sharan S. Agrawal

August 2023, 84 pages

This technical report is based on a dissertation submitted June 2023 by the author for the degree of Master of Philosophy (Advanced Computer Science) to the University of Cambridge, Darwin College.

DOI: 10.48456/tr-985

Abstract

As the world faces twinned crises of climate change and biodiversity loss, the need for integrated policy approaches addressing both is paramount. To help address this, a new agent-based model (ABM), the VDSK-B, was developed. Using Dasgupta’s review of the economics of biodiversity, it builds on the Dystopian Schumpeter meets Keynes (DSK) climate economics model to link together the climate, economy and biosphere. To our knowledge, this is the first ABM proposed that integrates all 3 key elements.

Existing ABM frameworks struggled with global policy design needs due to their inability to scale to planetary-sized models, and optimize model parameters at the large scales needed for policy design. A new ABM framework called SalVO was built using a formalism for ABMs that expressed agent updates as recursive applications of pure agent functions. This formalism differs from existing computational ABM models but is shown to be expressive enough to emulate a Turing complete language. SalVO is built on a JAX backend and designed to be scalable, vectorized, and optimizable. Employing hardware acceleration, tests showed it was more performant and more able to scale on a single machine than any existing ABM framework, such as FLAME (GPU).

Techniques for using backpropagation to create optimized policies differentiable, deterministic ABMs were further extended and implemented in SalVO. A novel protocol, GP-ABM, using William’s REINFORCE algorithm, was developed to optimize parameters in non-differentiable, stochastic ABMs. Both approaches are shown to be able to optimize ABMs for thousands of parameters, with backpropagation learning a highly non-trivial policy to move the centroid of a flock to a target location. This represents an innovation over current state-of-the-art techniques, such as Simulated Minimum Distance, which do not scale past fifty at most.

Finally, the VDSK-B model was implemented in SalVO, showing its capability of expressing highly complex ABMs. SalVO proved to be highly scalable, running a 5x bigger version of VDSK-B using just 4% of the time taken by the current open-source implementation, significantly strengthening its position as a preferred tool for large-scale ABM studies. While further work remains to be done on VDSK-B’s calibration and correctness, SalVO’s marriage of speed, scale and optimization has the potential to reshape how we approach, design, and apply agent-based models.

Full text

PDF (4.8 MB)

This report is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence.

BibTeX record

@TechReport{UCAM-CL-TR-985,
  author =	 {Agrawal, Sharan S.},
  title = 	 {{Scalable agent-based models for optimized policy design:
         	   applications to the economics of biodiversity and carbon}},
  year = 	 2023,
  month = 	 aug,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-985.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-985},
  number = 	 {UCAM-CL-TR-985}
}