The U.S. military experienced logistics challenges with land-locked Afghanistan, but one of the last times it faced actively contested logistics was with the German submarine wolfpacks in World War II. Operations research and systems analysis (ORSA) was born out of this era, and it has been rumored that ORSA analysts knew where the wolfpacks would patrol before the submarine captains were given their orders. Today, many are turning toward quantitative science again in the hope of finding ways to mitigate potential challenges while providing supplies to warfighters across contested regions. This focus is indeed warranted. Since World War II, mathematics has been exploited to make huge strides toward maximizing profit in commercial logistic enterprises. Many leaders look to artificial intelligence and machine learning (AI/ML) to bring about the next wave of innovation. However, merely copying successful commercial practices will leave supply chains vulnerable while wasting valuable resources chasing solutions before defining problems. It may go against conventional wisdom, but this article argues in favor of irrational, non-optimal, and unpredictable actions.
There has been an abundance of hope placed in the advancement of AI/ML, especially by those who are woefully unaware of how it works. In the simplest of terms, AI/ML needs decision-makers to optimize X by training the model using data from Y. This method of model creation can lead to driverless convoys and more effective preventative maintenance but can also fall short in addressing contested logistics. Where is the data from prior contested sustainment operations in similar conditions that can be used to train the model? One solution is creating synthetic data from simulations, but the AI/ML output may amplify any bias in the simulation and produce fictitious data, also known as hallucinations. War, and by extension contested logistics, should be an outlier, and therefore AI/ML has minimal training data to provide insights on how to optimally get supplies from point A to point B through contested routes. So, how did they do it in World War II?
Traditionally, a logistics routing problem is modeled as a network with nodes being source, demand, or transit points, while the connecting arcs convey information about the cost or risk associated with moving between nodes. Edsger W. Dijkstra’s algorithm is a well-known method that can quickly identify the path between any two nodes with the least cost. Again, this cost could be distance, money, or risk. Additional constraints, such as source nodes having limited supplies or demand nodes requiring minimal amounts, can be added. Optimization techniques such as linear, non-linear, and stochastic programming can help determine which supply routes carry the least cost within those constraints. Unfortunately, if the enemy has this information, they, too, can identify which routes the Army should take. In a contested environment, the Army would be best served by taking the less likely and potentially non-optimal route. Indeed, there must be a way to randomize the routes optimally. Enter game theory.
Game theory can potentially provide mixed strategies — a list of probabilities associated with the routes instructing how to use them to minimize the chance of interdiction. The advisory would also have an optimal strategy that maximizes the enemy’s chances of finding U.S. sustainment forces. For example, for each resupply, the Army would randomly pick one of three routes with the following likelihood: route 1, 50%; route 2, 25%; and route 3, 25%. This adds a layer of randomness to the strategy, but it assumes perfect information, and the adversary also knows the Army’s intentions. Game theoretic models can account for imperfect information and more complexity, but there’s the flaw of rationality in the end. Game theory relies on rational players playing for the strategies to be optimal. If the wolfpack commanders were more irrational, finding them would have been more challenging.
This is not to say technology and AI/ML cannot aid in contested logistics; it just means the Army needs to think differently than its commercial counterparts. Systems using advanced algorithms can pick up on deviations from normal expenditures much faster than humans and provide courses of action from the warfighter to the factory to aid decision-making. Instead of optimizing on cost, the Army optimizes its resiliency to disruptions within acceptable and quantifiable risk. It can also use AI/ML to assist it in being as random as possible in supply routes if the model is optimized to increase survivability and not efficiency. However, with all this in mind, Soldiers must still train using traditional planning factors should the enemy’s disruption affect the physical environment and the cyber network that underpins connectivity.
The Army should not rely on only commercial industry practices to help it prepare for contested logistics. Hurricane season may produce disrupted logistics, but hurricane season is fairly predictable, and the weather does not actively seek and pursue to prolong the disruption. Using AI/ML to overcome contested logistics will only be as successful as the quality of experience (data) fed into the model and the output we train it to achieve. It would be foolish to optimize supply lines with predictable routes and razor-thin margins. The Army needs to be as irrational as possible until its rational options are secured from enemy influence. There is a way forward where data, data science, and mathematics provide useful insight into navigating contested logistics, but it will take adopting a perspective far different from peacetime commercial operations.
Dr. William T. Smith has over 20 years of experience with operations research and logistics. He holds graduate degrees in both mathematics and industrial engineering. He currently teaches future operations research analysts at Army Sustainment University, Fort Gregg-Adams, Virginia.
This article was published in the Winter 2024 issue of Army Sustainment.