Stochastic Optimization, Mathematical Programming, Machine Learning, Unsupervised Learning, Queueing Theory, Strategic Queueing, Markov Decision Processes.
Data-Driven Decision Making, Healthcare Operations, Decision Support Systems, Service Operations, On-Demand Platforms.
The performance of Emergency Medical Services (EMS) operations depends critically on the dispatch of ambulances in response to emergency calls. Good decisions can literally save lives in the short term, while retaining the ability to respond to calls in the near future—conversely, bad decisions can have dire consequences. We partner with University of Pittsburgh Medical Center to develop a new data-driven approach to EMS dispatch, combining principles from machine learning and optimization. We propose a novel data-driven stochastic optimization approach based on conditional clustering. First, for every emergency call, we construct a set of similar calls, using historical data—thus defining a large set of possible future scenarios. Second, we cluster these scenarios using a new distance metric that captures the spatial-temporal patterns of urban operations. Third, we use the resulting clustering outputs to build a set of representative scenarios, which we integrate into a multi-stage stochastic optimization framework to support EMS dispatching decisions. Initial results suggest that the proposed approach can reduce the percentage of late responses by as much as 30%, as compared to a baseline policy that systematically dispatches the closest ambulance.
We study an on-demand platform's delay information disclosure policy when the platform serves two classes of users---consumers and providers---who seek matches to each other using the platform. We model the platform as maximizing the average rate at which these users are successfully matched by choosing one of three information regimes — occupancy (disclosing the current system occupancy to both user classes), and two asymmetric information regimes (disclosing no information to one user class and occupancy information to the other). Arriving users are strategic and decide whether to join the system or not based on the delay information that the platform provides. In our base model, we consider users of each class as being either patient (will wait to be matched) or impatient (will join only if they expect to be matched immediately). We find that depending on the parameter setting, any of the three information disclosure policies could emerge as optimal; however, the optimal policy has a complicated dependence on the parameters. We analytically establish sufficient conditions driving these decisions. We also examine a special case of an imbalanced market where one user type is relatively abundant and characterize the platform’s optimal information regime in this case. Again, all three disclosure policies may emerge in equilibrium. We then numerically examine the impact of the platform's choice of information regime on users' welfare and find that the platform's choice also maximizes the welfare of both user classes when this choice is to disclose occupancy information to both classes. We extend our base model to study how the platform's information regime choice changes when user patience levels are more heterogeneous.
Security screening systems aim to identify malevolent people and illicit goods. But screening operations may also result in long wait times at checkpoints. Selecting appropriate screening procedures thus creates a trade-off between efficiency and risk. This is complicated by the heterogeneity of screening jobs (which pose various threat levels) and the strategic behaviors of human agents (who may renege prior to screening if perceived risk levels are too high). This paper applies a speed-quality trade-off perspective to security operations. It extends the speed-quality trade-off literature to a multi-class setting with heterogeneous and strategic agents. This paper formulates continuous-time infinite-horizon Markov decision processes to optimize service rates in an M/M/1 queue with heterogeneous jobs, as a function of observed queue lengths and a threat level estimate for each job. We propose an extension to capture endogenous strategic behaviors of heterogeneous agents, given information asymmetries between agents and the screening operator. We find that the optimal policy exhibits a double threshold behavior: the shorter the queue length and/or the larger the risk, the stricter the screening. Leveraging job-level risk information can reduce expected costs by up to 6-7%, as compared to single-class decision-making schemes. Moreover, anticipating agents’ strategic behaviors results in more intensive screening in an attempt to force malevolent agents to renege. Slower screening mitigates expected risks and may also, surprisingly, reduce expected queue lengths.