TBM Gym: A toolkit for developing and comparing terrestrial biosphere models and non-convex optimization algorithms
Abstract
Terrestrial biosphere models (TBMs) utilize hybrid statistical-physical architectures containing non-differentiable functions, high-dimensional parameter spaces, and multiple outputs. Prior distributions may be specified over model parameters, utilizing sampling strategies to estimate uncertainty. TBMs are computationally expensive and difficult to optimize, as objectives may be non-convex and non-smooth, i.e., non-differentiable, or the gradient or Jacobian of an objective may violate assumptions of convexity or Lipschitz continuity. In this context, brute-force search is expected to be an NP-hard problem with greater than polynomial time complexity. As in deep neural networks, non-convexity may arise even from the composition of convex functions. This inhibits direct Hessian-based optimization (e.g., Newton's Method) but may allow quasi-Newton methods such as L-BFGS and SGD in some cases. Meanwhile, machine learning emulators have demonstrated utility for efficiently approximating optimality. Due to their complexity, TBMs provide an ideal arena for benchmarking and comparing single- and multi-objective optimization algorithms, where rapid convergence is desired.
Inspired by OpenAI Gym, we provide a collection of test problems, called environments, for developing and comparing TBMs and optimization algorithms. Environments consist of preconfigured TBM classes and methods to facilitate model parameterization, execution, and robust loss estimation. Minimal data acquisition, visualization, and helper methods are provided, including an interface to Google Earth Engine and machine learning emulators. Baseline optimization algorithms include particle swarm optimization, differential evolution, evolution strategies, Bayesian optimization, MCMC, reinforcement learning, and multi-objective evolutionary algorithms. We leverage state-of-the-art Python facilities, including GPU-accelerated deep learning and probabilistic programming to provide these capabilities. We demonstrate TBM Gym on a forest biogeochemistry model, Sortie-PPA. Unlike more elaborate tools such as PEcAn, TBM Gym is designed to be lightweight, minimal, and portable, fully operable within Jupyter Lab. Minimal R classes and functions may also be provided.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.B41N2915E
- Keywords:
-
- 0414 Biogeochemical cycles;
- processes;
- and modeling;
- BIOGEOSCIENCESDE: 0439 Ecosystems;
- structure and dynamics;
- BIOGEOSCIENCESDE: 0466 Modeling;
- BIOGEOSCIENCESDE: 1910 Data assimilation;
- integration and fusion;
- INFORMATICS