Towards a System Dynamics Framework for Human–Machine Learning Decisions: A Case Study of New York Citi Bike

Ganesh Sankaran, Marco A. Palomino, Martin Knahl, Guido Siestrup

The growing number of algorithmic decision-making environments, which blend machine and bounded human rationality, strengthen the need for a holistic performance assessment of such systems. Indeed, this combination amplifies the risk of local rationality, necessitating a robust evaluation framework. We propose a novel simulation-based model to quantify algorithmic interventions within organisational contexts, combining causal modelling and data science algorithms. To test our framework’s viability, we present a case study based on a bike-share system focusing on inventory balancing through crowdsourced user actions. Utilising New York’s Citi Bike service data, we highlight the frequent misalignment between incentives and their necessity. Our model examines the interaction dynamics between user and service provider rule-driven responses and algorithms predicting flow rates. This examination demonstrates why understanding these dynamics is essential for devising effective incentive policies. The study showcases how sophisticated machine learning models, with the ability to forecast underlying market demands unconstrained by historical supply issues, can cause imbalances that induce user behaviour, potentially spoiling plans without timely interventions. Our approach allows problems to surface during the design phase, potentially avoiding costly deployment errors in the joint performance of human and AI decision-makers.