Paper: arXiv 2307.01085

Authors: Arnau Quera-Bofarull, Joel Dyer, Anisoara Calinescu, Michael Wooldridge

Abstract

Agent-based models (ABMs) are a promising approach to modelling and reasoning about complex systems, yet their application in practice is impeded by their complexity, discrete nature, and the difficulty of performing parameter inference and optimisation tasks. This in turn has sparked interest in the construction of differentiable ABMs as a strategy for combatting these difficulties, yet a number of challenges remain. In this paper, we discuss and present experiments that highlight some of these challenges, along with potential solutions.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper features advanced mathematical concepts like automatic differentiation and gradient estimation strategies, but the experiments are conceptual simulations (e.g., Brock & Hommes model) and lack real-world financial data, making it more theoretical than backtest-ready.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Challenges of Calibrating<br>Differentiable ABMs"] --> B["Methodology: Experiments on<br>Gradient Estimation & Training"]
  B --> C["Inputs: Financial Systems<br>Agent Behaviors"]
  C --> D{"Computational Process:<br>Calibration & Optimization"}
  D --> E["Key Findings:<br>Gradient Instability"]
  D --> F["Key Findings:<br>Simulation-Model Mismatch"]
  E --> G["Outcome: Potential Solutions<br>for Robust ABM Calibration"]
  F --> G

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a differentiating strategy that alternates between forward-mode AD (FMAD) and reverse-mode AD (RMAD) when calibrating agent-based models.

Source sentence
To address this, in subsection 3.2 we discuss a differentiating strategy that alternates between FMAD and RMAD when calibrating ABMs, and we apply it to an epidemiological simulation involving 8 million agents. 2.3.

Assumes 3

  • The loss function ℓ(y,θ) is chosen to be a differentiable function of the model parameters θ. model
    Source sentence
    (6) Further details are provided in Appendix B. Thus, provided ℓis chosen to be a differentiable function of θ, this enables us to employ gradient-based approaches to minimising the objective (5) that exploit the reparameterisation trick.
  • Synthetic data y are generated from the Brock & Hommes model at known parameter values for calibration experiments. data
    Source sentence
    Fixing g1 = b1 = b4 = 0 and g4 = 1.01, we consider the task of calibrating parameters g2, g3, b2, b3 given syn-thetic data y = (y1, . . . , yT ) generated from the model at (g2, g3, b2, b3) = (0.9, 0.9, 0.2, −0.2) with T = 100.
  • The maximum mean discrepancy (MMD) loss used for calibration is differentiable and deterministic when a Gaussian kernel is employed. model
    Source sentence
    We follow Cherief-Abdellatif & Alquier (2020) and target the generalised posterior (4) given by the choice ℓ(y, θ) = MMD2(PT , Pθ), (7) where MMD2(PT , Pθ) is the maximum mean discrep-ancy between the empirical measure of returns PT = (y1, . . . , yT ) and the distribution Pθ of returns implied by the simulator at parameters θ.

Data

  • Synthetic data from the Brock & Hommes model synthetic unknown
    Source sentence
    Fixing g1 = b1 = b4 = 0 and g4 = 1.01, we consider the task of calibrating parameters g2, g3, b2, b3 given syn-thetic data y = (y1, . . . , yT ) generated from the model at (g2, g3, b2, b3) = (0.9, 0.9, 0.2, −0.2) with T = 100.
  • Synthetic time-series of daily infections for the JUNE model synthetic unknown
    Source sentence
    We generate a synthetic time-series of daily infec-tions for 50 days using some assumed parameters that we aim to recover through our calibration process.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Using a finite gradient horizon (H=0) dramatically improves the gradient‑assisted training of the variational posterior for the Brock & Hommes model. performance
    Source sentence
    In this experiment, the best performance was observed while using a gradient horizon of H = 0.
  • With the vanilla pathwise derivative, the objective decreases slowly and offers no improvement over the score‑based gradient. negative
    Source sentence
    This can be seen in Figure 1, in which the objective function decreases slowly with the number of epochs when trained with AdamW (Loshchilov & Hutter, 2017) and using the vanilla pathwise derivative (red curve).
  • The GVI calibration of the JUNE model converges after approximately 3,000 model evaluations, demonstrating simulation‑efficient calibration with gradient‑assisted methods. performance
    Source sentence
    The flow converges after approxi-mately 3,000 model evaluations, highlighting the potential for simulation-efficient calibration with gradient-assisted methods.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 9 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors)

Code

https://github.com/arnauqb/blackbirds — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.