The AI Evolution and the Challenge of Decision Automation

Warren Powell, co-founder, Optimal Dynamics, Professor, Princeton University

There is an unending stream of articles on the internet talking about the “AI Revolution,” describing how it is going to transform industry. This parallels the 1970s when “AI” and rapidly maturing computers were making the same claims… until it crashed with the disappointment of not meeting expectations.

We have been living in yet another AI wave, starting around 2005–2010. A recent article in The Economist: Technology Quarterly (June 13, 2020) quoted estimates by several consulting firms that AI will add $13-$16 trillion to the U.S. economy. Sundar Pichai of Google has described AI as “more profound than fire or electricity.”

Yet, reality is starting to settle in. CapGemini estimated that 1 percent of AI projects in logistics have been successful. The article in The Economist was titled “The Limits of AI.” It appears that we are in the early stages of another crash as AI, once again, does not meet these lofty expectations.

The problem, of course, is not with AI, but with the expectations.

AI represents a set of tools that automate information and decision processes, paralleling the use of robots in manufacturing. Robots represent the natural evolution of machines (which date back thousands of years) to machine tools (which became a major indicator of the strength of the manufacturing sector) to industrial robots.

Robots have been most visible in their use in automotive assembly, but their introduction by traditional manufacturers has been quite slow, starting with jobs like painting and welding which were unattractive to humans. Lofty expectations about putting workers out of jobs proved overblown, but steadily, incrementally, they are expanding the tasks that they can perform.

The most dramatic attempt to expand the use of robots has been by Tesla, where Elon Musk dreamed of manufacturing lines with 100 percent robots. This failed, with even Musk saying that “humans are underrated.” But this does not mean robots failed — it just means they did not meet Musk’s aspirational dream.

This has been the path with AI. Artificial intelligence in the 1970’s consisted primarily of rules: “If this condition, then do …”. The problem with rule-based AI is that it does not scale to complex problems where the condition has more than a few dimensions — the classic “curse of dimensionality.” This means that rules could not solve all the lofty ambitions of AI — they are clumsy, they do not scale, and they do not adapt…. but they are still incredibly useful, and widely used in virtually all information systems. Rule-based AI did not fail; it just failed to meet unrealistic expectations.

Almost all articles in the lay business literature on artificial intelligence today equate AI with neural networks, a technology that dates back to the 1950’s, but which exploded in the public imagination after 2005 when they emerged as a powerful tool for pattern recognition. Neural networks have proven to be exceptionally useful for recognizing images, speech, handwriting and patterns in text. However, people are starting to project these successes to problems for which they are simply not designed, paralleling the claims for rule-based AI in the 1970s.

Neural networks have been very successful for pattern recognition problems, which are large, complex, relatively unstructured problems, where (and this is key) the correct “label” for an input pattern is fixed (we would say deterministic). An image of a chair is always a chair. Neural networks are high-dimensional functions, often with hundreds of thousands to millions of parameters, which means that they can be trained to fit almost any shape. The price of this flexibility, however, is that they have to be trained on very large datasets. These labeled datasets, however, are not free, and they do not come from a computer. Labels require people, as highlighted by the New York Times.

Problems in freight transportation and supply chains, on the other hand, exhibit considerable structure, and a lot of noise. While pattern recognition technologies do have applications in this sector (interpreting emails, faxed documents, text translation), logistics is all about managing resources: drivers, tractors, trailers, and inventories.

Managing resources requires making decisions. The operations research community has developed powerful optimization solvers that provide perfect answers, but only if you have perfect data (and forecasts). The machine learning community developed their own decision tools under the heading of “reinforcement learning.” RL emerged when it successfully solved the Chinese game of Go, and has suddenly become the new hammer for solving all problems, including the highly complex domain of “dynamic resource allocation” (see the BBC article here). However, RL has been limited to small problems such as video games, controlling robots, and other “single entity” problems and would never scale to high-dimensional resource allocation problems.

I claim that a neural network can never be trained to dispatch a fleet of trucks or manage a supply chain, although it may be used as a tool within a larger system. For this problem setting, neural networks exhibit two features that will work terribly in this problem setting: they impose no structure, and their millions of parameters will just fit the noise that is found throughout problems in logistics.

This is illustrated using a simple newsvendor problem, which is fundamental to any resource allocation problem where we have to supply resources (“newspapers”) to meet an uncertain demand. If we fix the level of resources, the profit depends on sales, which is random, producing the random scatter of observations in the figure to the right. If we average the profits over all the possible demands, we get the smooth red line (which we do not know in practice). If we use observed profits to fit a neural network, we get the wavy brown line, which is our best estimate of the red line (pretty terrible!). The problem is that neural networks are just too flexible, and do not take advantage of the known structure of the problem.

So, is AI doomed to fail in complex resource allocation problems such as those that arise in freight transportation and logistics? Of course not. We first have to move past the tendency to equate AI and neural networks. Neural networks will have their uses, but primarily for the problems where they have already been successful: pattern recognition.

Resource allocation problems require tools that have been developed that exploit their natural structure. Powerful algorithms for for designing networks and scheduling vehicles have been in use for decades, building on advances from the optimization community (which went through its own boom and bust in the 1990s). But these tools only work when everything is known.

More problematically has been the development of tools for making decisions over time in the presence of different sources of uncertainty. I refer to the study of sequential decision problems as “decision analytics” which builds on the tools of machine learning and a substantial literature known as “stochastic optimization.” Sadly, most of these tools come across as complex and computationally intractable, which is a problem in an industry that prizes simplicity and transparency.

Building on my career of solving problems in freight transportation and logistics, I finally cracked the code for optimizing these complex problems under uncertainty. Instead of finding optimal decisions (as is done with deterministic problems), we have to find effective “policies” which are methods for making decisions given what we know at the time. Designing policies requires understanding the structure of the problem, and the types of uncertainties that we have to accommodate. I will return to designing policies in a future blog. For now, what is important is that you have to customize policies to the problem, just as Tesla has to customize robots to a specific task.

So what are we trying to accomplish with our AI tools? Most of the time we are just trying to help people make better decisions. Increasingly, however, the real goal is to eliminate the rooms full of people making decisions. In other words, automation. People are expensive, slow, and do not adapt quickly to changing conditions. At the same time they are unparalleled at human communication (phone calls, emails) and handling complex situations (such as negotiating with drivers and warehouse managers).

Ultimately our goal with AI in freight transportation and logistics is the same as Elon Musk’s goal with his robots on his assembly lines. We are not going to achieve this with any specific analytic tool such as neural networks (the current wave), deterministic optimization (as was claimed in the 1990s) or the newest wave, reinforcement learning (as claimed in the BBC article). The question is: do we have all the tools we need in our toolbox?

Today (and it took me a long time to be able to say this) I think we have the analytic toolbox to tackle the complex decision problems that arise in transportation and logistics, although refinements will continue to be needed. This claim is based on the unified framework that I designed a few years ago, and which is summarized by the material at jungle.princeton.edu. However, there is more to the problem than just analytic tools.

The biggest barrier to automation in freight transportation and logistics is inserting advanced analytics into processes that have evolved with people in the loop. This means that we still have people talking on phones, reading emails, and receiving faxes (which stunningly have not vanished). Easily the most prominent use of neural networks in this sector is voice recognition and natural language processing, but considerable work is still required to refine these tools.

The path to full automation will be a long one, just as it has been with the automotive industry. Elon Musk learned the hard way that there are some jobs that robots simply cannot handle (yet). One task where robots struggled at Tesla was installing the complex wiring harness. One way to solve this is to build more sophisticated robots that can handle wiring. The other, which Musk is pursuing, is to get rid of wiring. This is a lesson that we need to take to heart.

Imagine if we had tried to automate dispatching taxis without changing the process. We would use voice recognition to handle phone calls from customers (which might use neural networks for natural language processing), and then design a system for communicating to cab drivers (more NLP). Of course, this system would have to handle concerns from the cab drivers, in the myriad accents that taxi drivers are famous for.

Now compare to what Uber did. All information is communicated through the Uber’s smartphone app. Trips are offered to drivers who can decline, which eliminates the occasionally complicated negotiations with drivers. Specialized analytics are used to choose which driver to assign, and what price to offer both the driver and the rider. Both of these are examples of “policies” designed specifically for the problem at hand.

As of this writing, there are efforts to re-engineer the freight process by Uber Freight and Convoy, but the jury is still out on these initiatives (as of now both depend on investment funding). The vast majority of the massive freight system in the U.S. is still a hybrid of humans helped by computers, where computers are primarily communicating and storing data with minimal intelligence. However, it is an evolving process, just as automation has evolved in automotive assembly.

Automating the information side of a physical system such as freight transportation and logistics means we have to think about the flow of information and decisions. When automating a manufacturing line, the movement of parts is fixed by the design of the assembly line, which is comparable to the flow of information. Robots use sensors to understand the physical system (comparable to recognizing the arrival of data or recognizing a verbal request) while the actions robots take to assemble components are comparable to decisions in our logistic system.

The flow of information in a logistics system is not as clean as an assembly line, but it is moving in this direction. Orders have been arriving electronically from shippers for many years. I have seen estimates that 80 percent of truck drivers have smartphones with apps that can be used for driver dispatch. Digital freight marketplaces are growing. Communication is a work in progress, but it is one where the industry has come a long way. I anticipate that the next frontier for automating communication is with the warehouses.

The next wave in automation for logistics is decisions: pricing, buying and selling assets, accepting loads, assigning drivers, managing inventories, routing trucks, scheduling machines, bidding, advertising, and so on. The transition from learning from data (machine learning) to making decisions is a big one. The toolboxes for machine learning are mature and widely available, with universities turning out thousands each year who know how to use these tools. By contrast, there is almost no general purpose commercial software for making decisions under uncertainty, primarily because of the incredible variety of different problem settings. In fact, there are over 15 distinct academic communities that work on making decisions under uncertainty — all with their own languages and tools. Years ago I began calling this the “jungle of stochastic optimization.”

A major breakthrough in my own work was the realization that all of these fields could be described by a single mathematical framework which involves finding the best policy for making decisions over time. Virtually every book offers its own way of making decisions, but I found that these could be organized into four major classes of policies. When faced with the rich array of decisions that arise in freight transportation and logistics, we are simply faced with identifying the class of policy that is best suited for each decision. I will return to the problem of choosing the best policy, but at this point I will say only that it requires balancing solution quality (minimizing costs, maximizing profits, …) along with simplicity, computational tractability, robustness, data requirements, and transparency.

At this point, the path to automation for controlling logistics systems is starting to become clear, but there is always work to do. The key is to identify all the decisions that have to be made, which parallels all the steps of a manufacturing process. Decisions require information, so we need to focus on how information is collected, communicated and stored. Some information is simply not known when a decision is being made, so this is where we turn to machine learning. Automating the information and decision process requires automating information collection, communication, estimation, and finally, the decisions.

You cannot build a house with a hammer. If you have the right toolbox, and understand how a house works, then you will succeed. But automating complex processes such as freight transportation, supply chains and complex resource allocation problems will be evolutionary, not revolutionary.

Optimal Dynamics is a New York City-based startup that has raised $4.4M to date in order to automate and optimize the logistics industry through the use of High-Dimensional Artificial Intelligence. To find out more please visit: www.optimaldynamics.com