Machine Learning and Optimization: The Art and Science of Dispatching
Intricacies of Dispatching
Dispatching operations can become complicated and time consuming when you are an NEMT provider who offers several forms of transportation services (Ambulatory, Cane, Wheelchair, Stretcher) using multiple vehicle form factors (sedans, small vans with wheelchair lifts, large vans for behavioral health), operated by your drivers and/or independent contractors. Running this kind of an operation involves juggling various parameters such as vehicle capacities, appointment thresholds, travel time, traffic congestion, profit and cost considerations. Goals include making the most of the vehicles available, lowering deadhead miles, meeting promised service levels, and running a profitable operation. As the size of transportation operations increases, the dispatching task grows exponentially in complexity. Consequently, we find that dispatchers take upward of many hours to finalize dispatching plans for the day. As the day unfolds, plans may need to be changed if there is a vehicle breakdown, driver no-show, trip cancellation, last minute trip request, will-call activation, unexpected wait times, etc. Mastering dispatch takes several years on the job. In the process, dispatchers develop their own thumb rules to manage different scenarios.
Science of Dispatching
In past decades, computing power has increased multifold, and state of the art optimization algorithms have made their way into dispatching software applications. Optimization algorithms aim to balance conflicting business conditions and generate a dispatch plan that maximizes coverage, vehicle usage, and profit while minimizing deadhead miles and drive time, and ensuring that trip service level commitments are met. These algorithms take only a few minutes to run but they save dispatchers several hours of manual dispatch plan creation. A major drawback of these algorithms is that they treat business constraints as fixed for a run. Some constraints may be treated as “soft” but optimization treats business parameters as deterministic so it can provide a solution within a reasonable amount of time. Optimization is what brings the best of SCIENCE to the process of dispatching.
Art of Dispatching
When dispatchers use a plan generated by optimization algorithms, they may tweak it, based on their knowledge and experience. These adjustments may include planning a trip on a route even if that means the driver reaches beyond the allowed late arrival threshold configured for the optimization algorithm. A dispatcher might be aware that the rider needs extra time to get to the vehicle from their apartment, and therefore the driver can afford to reach a few minutes later. In some situations, the dispatcher may change the plan, so the driver reaches earlier than the early arrival threshold configured for optimization engine. This may be because the pickup location has parking restrictions that add potential delay. Making these adjustments to dispatch plans requires knowledge of business operations, and that cannot be codified in optimization algorithm settings. This is the ART of dispatching that requires intuition developed through years of dispatch work experience. It’s a skill that is not easily transferred neither can it be integrated into a fixed set of static rules.
Learn As You Go
In the last ten years, a new class of algorithms called Machine Learning (ML) has found use in solving various categories of business problems. ML algorithms differ from Optimization algorithms in one vital aspect; they do not rely on fixed set of rules and constraints. Instead, they use data from past business scenarios to infer the mechanisms that drive decisions. As more data becomes available, these algorithms are able to fine tune their internal parameters, and recommend smarter solutions. Every correction made by a user to the output feeds back into the algorithm and helps in improving a future result. One major drawback with ML algorithms is that their output relies heavily on the historical data used to “train” them. ML algorithms are able to codify business rules without predefining them explicitly, and this presents interesting opportunities as related to dispatch operations. Planning a specific trip assignment where the driver arrives outside of the planned threshold, or selecting a particular driver for a trip on the basis of positive rider feedback are both data-driven decisions best handled using ML algorithms.
Blending Art and Science
Combining the science (Optimization) and the art (data-driven ML) is the holy grail of dispatching. At WellRyde, we are deeply invested in utilizing vanguard technology to make transportation operations both efficient and profitable. The WellRyde platform offers a dispatch solution that combines best in class optimization algorithms with data-driven ML algorithms that learn from past dispatcher actions. Dispatchers configure business rules and constraints in order for the optimization algorithms to create an initial dispatch plan. After the plans are generated, the ML algorithms make adjustments based on similar trips in the past. Dispatchers can review the plans to make further tweaks, if required. These modifications feed back into the ML algorithm for recalibrating the next plan. The solution goes further by utilizing the optimization algorithm to identify opportunities for inserting on-demand trips that may be added at the last moment, thus saving time for the dispatcher. Any adjustments made are also fed back into ML algorithms to enable better decision making in future. In case of repeat trips, ML algorithms scrutinize their profiles, automatically assigning them to routes, based on past assignment and adjustments. Any on-demand trips that come in are handled by optimization algorithms to identify the best route to fit them in, based on predefined business rules.