Reinforcement learning to develop policies for fair and productive employment: A case study on wage theft within the day-laborer community

This paper applies a reinforcement learning (RL) approach (batch Q-learning) to solve decision making problems toward the development of policies for fair and productive work for laborers in precarious employment situations. We present both single-agent and multi-agent settings. The first formulation more closely resembles the limited agency available to laborers today and the second is presented to address the research question of how to develop policies that allow both laborers and employers participate in employment decisions and to respond to unfair work conditions. The single agent formulation confirms a policy often observed in practice where day laborers take jobs with the risk of wage theft and endure the outcome because the likelihood of achieving justice is low and the laborer typically still receives a fraction of their wages. We demonstrate that the two-agent formulation allows the policy to encompass decisions by both laborers and employers. Within this decision-making dynamic, we illustrate through sensitivity analysis that under modest increases in the likelihood of a successful outcome of reporting, laborers learn to report theft and employers learn not to steal. We use the complexity of the case study examined to motivate a more general formulation based on the generalized semi-Markov process that allows the method to incorporate more detailed system dynamics that, in turn, allow for more precise policies to be formulated and determined. We discuss the implications of both the policies determined in the case study and the potential of the generalized semi-Markov reinforcement learning formulation.Author summary: This study continues previous translational research at the intersection of behavioral and decision science toward the larger goal of designing dynamic operational policies for complex sociological systems. While this paper has a methodological orientation, we situate a demonstration of our evidence-based model in the problem domain of “decent work” as envisioned by the United Nations Sustainable Development Goal 8. Here, we apply and extend reinforcement learning to observe the system, periodically try new policy ideas, and adapt the deployed policy based on feedback from the system under the new policy candidate. We apply our approach to the societal challenge of disrupting wage theft experienced by day laborers. We use this case study to recognize the benefit of adapting our approach to complex systems by incorporating a more generalized stochastic model.