Evolutionary model discovery of causal factors behind the socio-agricultural behavior of the ancestral Pueblo

Chathika Gunaratne, Iván Garibay, Nguyen Dang

Abstract: Agent-based modeling of artificial societies offers a platform to test human-interpretable, causal explanations of human behavior that generate society-scale phenomena. However, parameter calibration is insufficient to conduct an adequate data-driven exploration of the importance of causal factors that constitute agent rules, resulting in models with limited causal accuracy and robustness. We introduce evolutionary model discovery, a framework that combines genetic programming and random forest regression to evaluate the importance of a set of causal factors hypothesized to affect the individual's decision-making process. We investigated the farm plot seeking behavior of the ancestral Pueblo of the Long House Valley simulated in the Artificial Anasazi model our proposed framework. We evaluated the importance of causal factors not considered in the original model that we hypothesized to have affected the decision-making process. Contrary to the original model, where closeness was the sole factor driving farm plot selection, selection of higher quality land and desire for social presence are shown to be more important. In fact, model performance is improved when agents select farm plots further away from their failed farm plot. Farm selection strategies designed using these insights into the socio-agricultural behavior of the ancestral Pueblo significantly improved the model's accuracy and robustness.

Introduction

Agent-Based Modeling (ABM) serves as a critical methodology for simulating complex social dynamics, such as the historical trajectory of ancient civilizations. A persistent challenge in ABM research is the "black-box" problem, where researchers often pre-define decision rules—such as agents moving to the nearest available land—and focus solely on parameter calibration. This approach assumes the underlying behavioral logic is correct, which may lead to models that lack causal accuracy. To address this, we present Evolutionary Model Discovery (EMD), a framework designed to systematically explore and identify the causal factors driving observed social phenomena.

Methodology

The EMD framework employs a two-stage process to move beyond simple parameter tuning toward "white-box" rule discovery:

  1. Genetic Programming (GP): Rules are "evolved" using primitives (e.g., distance, soil quality) and mathematical operators. The GP selects for rules that best replicate historical population data from the Long House Valley (800–1400 AD).
  2. Random Forest Regression: A regressor is trained on GP results to quantify "feature importance," identifying which factors most significantly impact model fitness, measured by Root Mean Square Error (RMSE).

The study utilizes the Artificial Anasazi model, incorporating archaeological data on household counts, water sources, and soil dryness (Adjusted Palmer Drought Severity Index).

Results and Discussion

Our analysis reveals that the original assumption of proximity-based decision-making is insufficient. As shown in Figure 5, minimizing distance ($F_{Dist}$) was among the least critical factors for model performance. Instead, Land Quality ($F_{Qual}$) and Social Presence ($F_{Soc}$) emerged as the primary drivers of accurate simulations. The model also performed better when agents were permitted to move further from failed plots ($F_{Mig}$), rather than remaining nearby.

Figure 5: Impact of various causal factors on model RMSE. Note the significant error reduction associated with Land Quality (F_Qual) and Social Presence (F_Soc) compared to Distance (F_Dist).

Furthermore, the availability of information significantly influenced model accuracy. Comparing different information scenarios, agents with "Full Information" ($S_{All}$) about the valley's resources achieved the lowest RMSE, suggesting the existence of robust communication networks within ancestral Pueblo communities.

Figure 3: Comparison of Root Mean Squared Error (RMSE) across different information scenarios. The S_All scenario, representing full information access, yields the most accurate model fit.

In terms of performance, 15 out of 20 genetic programming runs produced rules with lower RMSE than the best results previously recorded in literature. Specifically, rules prioritizing land quality ($argmax(F_{Qual})$) provided a significantly better fit than the original distance-based logic. These discovered rules also demonstrated superior robustness; while the original model's performance varied significantly with starting parameters, the evolved rules maintained consistent accuracy.

Conclusion

Evolutionary Model Discovery successfully identified that ancestral Pueblo farmers likely prioritized land quality and social community over mere proximity when selecting farm plots. While EMD is dependent on the initial set of hypothesized factors and is computationally intensive, it offers a powerful path forward for discovering the hidden drivers of both ancient and modern social behaviors. Future work will apply this framework to contemporary issues such as urban migration and digital information spread.

Summary updated

Abstract

Agent-based modeling of artificial societies offers a platform to test human-interpretable, causal explanations of human behavior that generate society-scale phenomena. However, parameter calibration is insufficient to conduct an adequate data-driven exploration of the importance of causal factors that constitute agent rules, resulting in models with limited causal accuracy and robustness. We introduce evolutionary model discovery, a framework that combines genetic programming and random forest regression to evaluate the importance of a set of causal factors hypothesized to affect the individual's decision-making process. We investigated the farm plot seeking behavior of the ancestral Pueblo of the Long House Valley simulated in the Artificial Anasazi model our proposed framework. We evaluated the importance of causal factors not considered in the original model that we hypothesized to have affected the decision-making process. Contrary to the original model, where closeness was the sole factor driving farm plot selection, selection of higher quality land and desire for social presence are shown to be more important. In fact, model performance is improved when agents select farm plots further away from their failed farm plot. Farm selection strategies designed using these insights into the socio-agricultural behavior of the ancestral Pueblo significantly improved the model's accuracy and robustness.
BibTeX
@misc{gunaratne2018evolutionary,
  author = {Chathika Gunaratne and Iván Garibay and Nguyen Dang},
  title = {{Evolutionary model discovery of causal factors behind the socio-agricultural behavior of the ancestral Pueblo}},
  year = {2018},
  eprint = {1802.00435},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.00435},
}
Figure