Agent Based Modeling for Human-Animal Behavior and Interaction (Ongoing)

Project Overview

This study developed a robust, reproducible agent-based modeling framework to characterize fine-scale movement patterns of free-ranging savanna elephants using high-resolution GPS collar data. The pipeline enables rigorous statistical fitting, simulation, and comparative visualization of stochastic movement models, supporting ecological hypothesis testing on resource selection, search efficiency, and seasonal behavioral plasticity.

Key Components and Methodology

1. Data Ingestion and Preprocessing
   Raw GPS fixes (latitude, longitude, timestamp) from deployed collars were imported and projected from WGS84 geographic coordinates to a local UTM-based flat-earth metric system (meters). Outliers were filtered using speed and acceleration thresholds; zero-displacement steps (e.g., sensor noise, resting) were excluded to isolate active locomotion.

2. Step Length and Turning Angle Extraction
   • Step lengths: Euclidean distances between consecutive validated fixes.
   • Turning angles: Computed as bearing differences with modular wrapping (±π) to resolve array broadcasting issues during vectorized operations.
     Seasonal stratification (wet vs. dry periods) was applied using rainfall proxies and fix timestamps.

3. Lévy Walk Model Fitting
   Power-law distributed step lengths were fitted via maximum likelihood estimation (MLE) under a truncated Pareto distribution:
   [ P(l) \propto l^{-\mu}, \quad l_{\min} \leq l \leq l_{\max} ]
   The exponent μ was optimized numerically using scipy.optimize to assess super-diffusive foraging signatures.

4. Correlated Random Walk (CRW) Implementation
   A persistent directional bias model was constructed with:
   • Step lengths drawn from an exponential distribution (fitted via MLE).
   • Turning angles sampled from a wrapped von Mises distribution:
     [ f(\kappa | \mu, \kappa) = \frac{e^{\kappa \cos(\kappa - \mu)}}{2\pi I_0(\kappa)} ]
     Concentration parameter κ was estimated per individual and sex. Vectorized simulation resolved prior broadcasting errors through explicit modulo handling.

5. Model Simulation and Validation
   For each observed trajectory, 1,000 synthetic paths were generated under Brownian motion, Lévy walk, and CRW parameterizations. Net displacement, path tortuosity, and area restricted search (ARS) metrics were compared via Kolmogorov–Smirnov tests and visual diagnostics.

6. Sex-Specific Comparative Analysis
   Behavioral divergence between bulls and matriarchal family groups was quantified. Visualizations included:
   • Kernel density estimates of turning angle distributions.
   • Superimposed real vs. simulated trajectories.
   • Quantile–quantile plots of step length distributions.

Technical Contributions

• Engineered a modular Python pipeline (pandas, numpy, geopandas, scipy, matplotlib) for end-to-end telemetry analysis: projection, filtering, feature extraction, model fitting, simulation, and visualization.
• Implemented statistically rigorous MLE fitting for Lévy and CRW parameters with confidence interval estimation.
• Resolved numerical stability issues in turning angle computation through robust modular arithmetic.
• Designed comparative diagnostic plots enabling model selection and ecological interpretation.
• Adhered to movement ecology best practices: outlier removal, temporal autocorrelation handling, and seasonal contextualization.

Acknowledgments

This work was conducted with the guidance from Prof. Roshan Godaliyadda, Prof. Parakrama Ekanayake from the Department of Electrical and Electronic Engineering, University of Peradeniya, Sri Lanka, Dr. Vandercone from the Department of Zoology, University of Peradeniya.