We apply two AI methodologies to the needs of trucking. The first approach is what is known as statistical AI, or more popularly as machine learning. It is premised on the idea that a large volume of historical, current, and future data has patterns of importance. The software discovers those patterns, and uses them to predict key outcomes. It is called machine learning because greater experience improves the predictive power of the software.
The second approach is known as AI Planning. Planning does not necessarily require learning from experience (although that can help). Instead of focusing on learning from experience, AI planning is pursued by accurately describing the state of the world, the actions available, and goals we want to achieve. With that, the AI acts as a rational agent trying to achieve goals with allowed decisions and with an expectation of how the world will react to them.
The diagram above shows how our AI is architected to leverage both AI methods. The planning side is about producing logically superior ways to use a truck. It ends with what we call “bundles”, which are specific ways to schedule, consolidate, sequence, and therefore execute a combination of transport orders.
But a rationally good bundle still needs to be qualitatively attractive to the person who decides to accept it. These are dispatchers at a carrier deciding what what to do with their trucks. The fact that they are in a different company means they have mis-aligned incentives and will not want to disclose all their information about capacity, price sensitivity, and so forth. Also, behavioral economics shows that most people are not aware of their own true preferences and these are only revealed indirectly.
Machine learning is used to construct a specific tender, for each bundle, to each relevant carrier. The machine learning here comes in two types. One is reinforcement learning on the selection of what strategy to use for each bundle. The second type of machine learning is supervised learning during strategy generation itself. Strategies themselves are defined in more detail here.
All of these AI methods are applied “in the middle”, between the cargo order coming from the TMS and the tender given to the carrier. If TNX does its work right, the two sides don’t see the complexity of the AI, just the quality of its results.