Human decision-making and behavior play significant roles in the introduction, spread, recognition, reporting and containment of new, emerging or foreign diseases and pests. Detection and mitigation strategies against the introduction of disease are commonly termed “biosecurity”. While there are generally accepted biosecurity best management practices (BMPs) to support herd health in food animal production systems before, during and after a disease outbreak, an improved understanding of the leverage points and mediating factors of biosecurity BMP adoption at the producer level can inform the design of biosecurity risk management strategies. In this study, we developed a Partially Observed Stochastic Game (POSG) to test the role of information uncertainty, risk messaging and economic incentives on the producer behaviors pertaining to the adoption of biosecurity BMPs. The POSG is mounted online and more than 1000 Amazon Turks are invited to play the game containing 32 treatments over 5 rounds, leading to 1000x32x5 = 160,000 observed decisions. Earlier, about 100 undergraduate students played the game in a lab environment, generating 16,000 observations. In both online and lab environments, subjects received differentiated payments matched with their performance in the games. The gaming data (N=176,000 observations) are analyzed with a variety of supervised and unsupervised machine learning algorithms to learn likely BMP adoption strategies under different treatments. The best fitting unsupervised machine learning algorithms, derived from 10-fold validation by splitting the gaming data into training and test samples, are used to drive the behaviors of producer agents in a multi-level agent-based model (ABM). The ABM was initially developed to assess factors relevant to the spread of socioeconomically important diseases through livestock production chain networks. The model used a GIS-based spatial framework, with three important hog-producing U.S. states—North Carolina, Iowa, and Illinois—defining the study areas. Four types of agents existed in the model, these being (a) hog producers, (b) feed mills, (c) slaughter plants, and (d) auction houses. In the ABM developed relevant to porcine epidemic diarrhea virus, the movements of animals and feed deliveries were the pathways of disease transmission. Biosecurity practices, such as disinfecting livestock transportation vehicles or constructing shower-in / shower-out facilities, were designed to reduce disease transmission probabilities. The network links were associated with infection spread probabilities that depend on the biosecurity levels of the agents and the type of transfer. For a baseline experimental simulation, we ascribed perfectly rational behaviors to all agents in the ABM. Given the complex nature of POSG, derivation of Nash equilibria is an NP-hard (non-deterministic polynomial-time hard) problem. Instead we ascribed rational behaviors to the agents by assuming risk neutrality, perfect memory and reinforcement learning using expectation maximization and Q-learning algorithms. Further, unsupervised algorithms learned from the online gaming data were used to generate more than one million alternate experimental simulations in the ABM and identify leverage points for risk management interventions by ranking the relative importance of risk messaging, incentives and information uncertainty in minimizing the disease outbreak risks. Implications of combining big data games with ABMs for complex systems research will be discussed.
