MACHINE LEARNING, AI, AND ROBOTICS^*

Today, “artificial intelligence” (AI), “machine learning” (ML), and “robotics” are terms that many of us hear on a daily basis. While previously AI was considered mostly science fiction, such as portrayed in the 1984 classic movie The Terminator, today we interact with and hear about AI and robotics applications in our homes (Amazon Alexa), at work (automated assistants), or in public spaces (public transport networks). Inevitably, AI, ML, and robotics have a strong impact on digital supply networks (DSNs) today, which will only grow in the future.

In this chapter, we will cover the impact of advanced AI and ML algorithms as well as robotics applications on today’s and future DSNs. First we will provide an overview of AI, ML, and robotics. Then we will put these terms into context, explain the key terminologies, and provide insights on their application in DSNs. The first section focuses on the background and terminology of AI and ML, before we have a take a closer look at the different algorithms and the process of applying AI/ML to a DSN problem. The third section focusses on robotics and automation in DSNs. In the last section, we discuss several examples to visualize the applications, challenges, and potential benefits of their application in DSNs. We conclude the chapter with 5 + 1 rules for successful use of AI in DSNs.

Before we start to dive deeper in the topic, let’s revisit what triggered this development. While AI, ML, and robotics are not a new topic, only in recent years do we see value-added applications across domains within our daily lives outside of dedicated, special solutions in the military, factories, or universities. Two key enablers are the increased connectivity along with the wealth of available data, and the progress made with regard to available computing performance. Two widely referenced laws, Metcalfe’s law and Moore’s law, exemplify the development (see Figure 4.1). Metcalfe’s law describes the exponential effect of an increasing number of nodes in a communication network on the network’s impact. Moore’s law focusses on the high speed at which computing power increases by stating that the number of transistors on a dense integrated circuit doubles every 18 months.

FIGURE 4.1   Rapid Technological Development Benefits AI Solutions

In the previous chapter we highlighted how the availability of data is key for any AI and ML solutions, and by extension robotics applications as well. Technological advances directly impact the cost of utilizing AI, ML, or robotics and thus lower the barrier of implementation and increase the competitiveness of applications in a DSN business context (see Figure 4.2). Furthermore, AI, ML, and robotics enable the establishment of platforms and thus solutions that are scalable across whole DSNs and beyond.

FIGURE 4.2   Improved Technological Basis for AI Applications

AI, ML, and robotics are all attempts to automate certain processes. In the case of robotics, these processes involve at least partly a physical component, such as a welding robot in an automated assembly line. For AI and ML, the automation is focused on a cognitive level. Therefore, we can distinguish AI and ML as cognitive automation, and robotics as physical automation. The purpose of these automation attempts is very important: generally we try to automate tasks that are repetitive, strenuous, and/or dangerous (see Table 4.1). Repetitive includes the need for extreme technical precision, such as placing fiber tape precisely in an automated fiber placement (AFP) process in composites manufacturing to avoid wrinkles and other problems. While for physical automation the descriptors are more clearly defined, in cognitive automation they are more indirectly associated. In both cases, the intention is to free up the human operator to do what humans do best: problem solving, creativity, and innovation.

TABLE 4.1   Examples for cognitive and physical automation tasks

In most cases, the three descriptors (repetitive, strenuous, and dangerous) are not mutually exclusive. An example of a typical DSN application for such a multidimensional problem is self-driving trucks. Consider an LTL (less than truckload) scenario where goods are to be transported from A to B, with A and B both located within busy metropolitan areas. Automating the whole route is challenging given the complexity of navigating in a dense city environment (close to A and B), loading and unloading, and the need to interact with personnel at A and B, just to name a few. However, once the truck enters the highway, the task at hand is much less complex with regard to automation potential, and at the same time can be classified as repetitive (driving hours upon hours on one street at defined speed), strenuous (driver has to be alert and concentrated for long periods of time), and potentially dangerous (accidents). Therefore, in this case, utilizing self-driving trucking capabilities for the highway portion of the route would solve multiple issues and address a pressing problem of the trucking industry at the same time: finding qualified drivers that are willing to spend most of their time away from home.

This example, while highlighting the three descriptors, also touches upon an important aspect when thinking about AI, ML, and robotics: the degree of automation. In most cases, implementing cognitive or physical automation is not a black-or-white issue, but there is an optimal degree of automation for the specific application and environment. This is essential to keep in mind when developing an AI or robotics strategy for your company or DSN.

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

AI and ML are not new, despite the recent increase in media coverage. The first definition from the 1950s describing ML as the process of allowing computers (artificial systems) to solve problems without being specifically programmed to do so¹ is still valid today.² Since then, AI and ML have developed significantly and gone through several cycles of growth and decline, most notably the so-called AI winter from the end of the 1960s to the mid-1980s when AI was declared dead and funding almost completely ceased. Since then, especially with the emergence of neural networks, we have seen a continued growth in both AI algorithms and applications. Figure 4.3 illustrates the key developments in the AI field from the mid-twentieth century to today.

FIGURE 4.3   Historical Perspective on AI

The terms AI and ML are often used interchangeably. However, there is a distinct difference that is key to manage expectations and understand how to add value through AI/ML in a DSN context. It basically comes down to the notion of general AI versus specific AI. General AI refers to an automated system that is truly intelligent and able to adapt to different situation within a very broad scope. In its extreme form, it can develop new systems and thus evolve. Specific AI on the other hand is focused on a certain domain, process, or problem and applies ML algorithms to process data, learn, and derive insights without constant supervision. Figure 4.4 presents a comparison of the definitions of AI and machine learning, highlighting the key differences. In DSN and most industrial applications we focus mainly on specific AI and thus ML instead of a general AI.

FIGURE 4.4   Artificial Intelligence (AI) vs. Machine Learning (ML)

Source: Mills, 2018

Technically, ML is a subfield of AI that is based on statistics, mathematics, and visualization. Within ML we use a variety of different algorithms to derive insights from big data. Figure 4.5 depicts the relationships between AI, machine learning, and data analytics as well as different specific algorithms in a graphical model.

FIGURE 4.5   Relationship Between Key AI Terminology

Let’s take the example of one of the most hyped ML algorithms today, deep learning, to illustrate the hierarchy of terminology with an association (Figure 4.6). Imagining AI as a human, we can imagine AI being the head, ML the brain, and the specific ML algorithm, in this case deep learning, the neurons in the brain utilized to process the data and providing actionable insights.

FIGURE 4.6   Hierarchy of AI, ML, and Specific Algorithm (Deep Learning)

The overall objective of AI and ML is to discover patterns in big data and provide a prediction of future behavior of a system or process based on the identified, often complex patterns. These predictions can target a variety of different issues, including detecting problems or process inefficiencies, market trends, customer behaviors, environmental impacts, financial indicators, or quality issues. AI and ML are the foundation for many new concepts such as predictive maintenance (see Chapter 9), DSN planning (see Chapter 6), and dynamic fulfillment (see Chapter 10) just to name a few. The ability to analyze large data sets, identify hidden patterns, and derive meaningful, actionable insights from the predictions are a powerful tool that has applications in virtually every aspect of business within a DSN.

AI AND MACHINE LEARNING ALGORITHMS

The objective of this section is to provide readers with a basic foundation and understanding of AI and ML to enable informed decisions relevant for their DSN strategy. We do not aim to present a dive deep into the different algorithms and the process of applying ML itself. To do so, we first have a look at the different classes of ML algorithms before covering the principal process of applying ML. Machine learning algorithms can be clustered in three principle classes:

•   UNSUPERVISED LEARNING. No label available.

•   SUPERVISED LEARNING. Label available.

•   REINFORCEMENT LEARNING. Feedback provided.

Unsupervised ML distinguishes itself from supervised ML in that it is used to analyze data that is not labeled. Labeled in this context means there is no evaluation of the examples available to provide insights to the ML algorithm. In ML terms, there is no teacher to add context and insights that can be utilized by the algorithm through the training data. Unsupervised ML tries to identify clusters in big data and thus provide actionable insights. Supervised ML, on the other hand, is built on the premise that the correct label (aka insight) is provided by an expert teacher. A common example is a quality label associated with a certain product instance to augment its parameters, state, and process data. Reinforcement ML algorithms do not rely on labeled data; instead they rely on feedback on actions taken through cumulative rewards. Therefore, compared to supervised ML, less feedback is given, since not the proper action (label), but only an evaluation of the chosen action is given by an expert teacher. Figure 4.7 illustrates a selection of different ML algorithms and their association with supervised or unsupervised principles. We can observe that several algorithms can be used in both an unsupervised and supervised fashion depending on the application case and data available.

FIGURE 4.7   Overview of Supervised and Unsupervised ML Algorithms

Source: Wuest et al., 2016

In DSN we have applications for both supervised and unsupervised algorithms. When we dive deeper, certain DSN subdomains are inclined more toward either supervised or unsupervised algorithms—mainly dependent on the nature of the problem and the data picture available. For example, in manufacturing, most applications fall within the supervised ML algorithm realm, while in customer segmentation, we see mostly unsupervised methods in play. Therefore, we focus mainly on the process and applications of supervised ML in the remainder of this section. In Figure 4.7 an overview of selected, noncomprehensive ML algorithms is presented. We can observe that most algorithms in this schematic can be applied to both supervised and unsupervised ML problems. Similarly, reinforcement learning can utilize several of the depicted algorithms.

Another way we can structure ML algorithms is by the type of prediction we want to achieve. In general we can distinguish between classification and regression as well as clustering. Classification attempts to predict the class of a specific example in a discrete way. For example, associating a part with the class “acceptable” or “scrap” falls in this category. Regression on the other hand predicts a certain numerical value as an output. A prime example is the stock price of a company at a certain time. Clustering is similar to classification with the key difference that it is applied in an unsupervised fashion. Clustering algorithms find patterns in data and group similar examples. This is used to segment customers based in selected features to apply directed marketing tools.

The process of applying ML algorithms starts with an appropriate data set. Different algorithms have variating requirements toward the data set itself. This includes but is not limited to the size of the data set (big data), data quality, and the balance of examples (see Chapter 3). Covering all aspects of how the nature of the available data set informs the choice of a suitable ML algorithm is beyond the scope of this section. However, common issues that impact the choice of an appropriate ML algorithm within DSN are (1) the balance of the data set in terms of labels and (2) the ratio of examples and features.

In a nutshell, most ML algorithms perform better when the data set provides a balanced distribution of the different labels (e.g., good quality vs. bad quality labels). In a common application area within a DSN, the manufacturing shop floor, we often fail to provide a balance of labels for a simple reason that we try to minimize scrap or quality problems. Therefore, the resulting data sets emerging from a manufacturing setting regularly display a significant overrepresentation of “good” quality label with only very few “not good”/“bad” labels. This is in itself not problematic and actually desirable as we do not wish to produce large quantities of low-quality parts. However, it presents a problem for many ML algorithms in terms of the training data set. In a scenario where we have a data set with 99 “good” examples and 1 “not good,” the algorithm can easily achieve a 99 percent accuracy by classifying all parts as “good”—and 99 percent accuracy is considered excellent in most cases. However, the resulting prediction would be totally useless in providing valuable, actionable insights to improve our process. Another issue is the discrepancy of examples (e.g., products/parts produced) and features (e.g., process parameters, quality measures). Both of these issues can be addressed and mediated during the preprocessing stage (see Figure 4.8).

FIGURE 4.8   Generalized Machine Learning Application Process

We want to make it very clear that the available data and the preprocessing stage are crucial to a value-added application of AI/ML in DSN. Many ML experts agree that data preprocessing is the most important step in the process of applying ML. A good data set with well-defined features with a basic algorithm often yields better outcomes than a low-quality data set without careful preprocessing and the use of a sophisticated algorithm.

The basic machine learning application process (see Figure 4.8) starts with the data set and the preprocessing of that data set. In a next step, a suitable ML algorithm is chosen based on the data set, the application objective, and other constraints such as computational resources and expertise available. Once the ML algorithm is chosen, we use a training data set to train the ML model. The training data is often created by splitting the available data set into 70 percent/20 percent/10 percent sections, with the training data set portion being the largest with about 70 percent of the data available. The remaining data is split into an evaluation data set (20 percent) that is used to optimize parameters, and a 10 percent test data set to test the model’s performance. However, in practice we often see a split with a training set of about 80 percent of the data and a test set of about 20 percent. Training, evaluating, and tuning the ML model is a continuous process that includes several loops before the ML model is ready for deployment.

ROBOTICS AND AUTOMATION IN DSN

When we think of robotics, most of us either think of industrial automation solutions, such as automated moving assembly lines in the automotive or semiconductor industry, or advanced (humanoid) robots such as the Terminator or the viral videos posted by Boston Dynamics. However, self-driving cars, drones, automated vacuum cleaners, and many shop-floor systems are also robots. For this section, we will use one of Merriam-Webster’s definitions for robot: “a device that automatically performs complicated, often repetitive tasks.”

Given the breadth of DSN and the diversity of tasks and stakeholders DSNs encompass, we encounter various opportunities for value-added applications of automated systems and robotics. Similar to the application of cognitive automation (AI and ML), physical automation and robotics applications differ depending on the different levels of abstraction (see Figure 4.9).

FIGURE 4.9   Different Levels of AI, ML, and Robotics Applications in DSN

On the lower abstraction levels, on the manufacturing shop floors, in warehouses, and in logistics centers, we tend to see more robotic systems today. Nevertheless, the shape and form of the robots in use varies significantly. They range from automated systems that encompass whole rooms, such as welding robots used on the assembly line of car body production, to smaller collaborative robots supporting the operator in loading individual machining centers.

When thinking of robotics and automation in a DSN, it is advisable to first think about the task that might benefit from being automated. Again, prime value-adding application areas for automation and robotics have certain characteristics in common: tasks are repetitive, strenuous, and/or dangerous. Setting up and programming a robotic system is not trivial and requires expertise. Therefore, introducing robotics to automate certain tasks needs to be carefully planned and critically assessed. In DSN, there are selected areas that tend to benefit more from automation than others. Figure 4.10 depicts a selection of areas where robotics are most likely to add value to the operations. Logistics is a whole big area in itself, and significant progress is made in that space daily. The list is not comprehensive, and with the rapid progress in the robotics space, new applications emerge regularly.

FIGURE 4.10   General Application Areas for Robotic Systems in DSN

For truly autonomous robotics applications, a field where AI and robotics heavily intersect, applications in the DSN space are still rare. However, individual noteworthy applications just recently emerged, such as autonomous drones for surveillance for energy efficiency and/or leak monitoring. The next area where autonomous robotic systems will make a large impact on DSN is logistics, including warehousing and transportation. Autonomous trucks, “dark” warehouse systems (fully autonomous with no human present), and last mile delivery systems (e.g., via drones) are on the horizon and will be transformative for DSN.

However, when planning robotics applications today, a key aspect that we need to consider is the optimal degree of automation. Automation, if done right, tends to reduce the direct cost per part; however, with higher degrees of automation, the indirect cost increases significantly (see Figure 4.11). This increase in indirect cost can be associated with expensive maintenance, control, and programming of the robotic systems. In essence, the management of the introduced complexity (via the robotic system) consumes much of the savings in direct cost. Therefore, there is an optimal degree of automation for a part and/or production line that delivers the best performance. It is crucial for DSN to (1) be aware of this phenomenon and (2) not always assume that “more is better” when it comes to automation. In many cases, a semiautomated production outperforms a fully automated system—for example, when dealing with small batch sizes. The goal of every DSN must be to achieve an as flexible, high-quality, and cost-effective production as possible.

FIGURE 4.11   Optimal Degree of Automation (Degree of Automation vs. Cost per Part)

In many cases the degree of automation and the cost of implementation is not a linear development. There is a sharp increase in cost at a certain degree of automation that must be carefully assessed to see if value provided justifies this increase. An example where such a rapid surge of cost regularly occurs is automating the operations of a machine tool such as a five-axis milling center. In this case we can differentiate five degrees of automation (see Table 4.2). The first degree of automation in this case describes the operator manually loading, ejecting, and moving the part as well as having no push notification from the machine tool. Degree 5 is the other extreme where the machine loading, ejecting, and part transport is all fully automated and the machine is providing push notifications to the operator. In this case, we can observe a steep increase in cost after degree 3, when moving toward implementing an automated loading robot. The complexity of the task requires precision and task-specific programming that drives cost. Prior to that task, equipping the machine tool with a notification capability (degree 2) and automating the ejection of the part (degree 3) does not come with such a price tag, as the tasks are comparably less complex to achieve. However, they already provide real value for the operations. Justifying automation beyond such “low-hanging fruits” within a DSN is not trivial, however necessary it might become to remain competitive on the marketplace.

TABLE 4.2   Sharp Increase in Cost at a Certain Degree of Automation (Example of Automation Options for a Milling Center)

The degrees of automation and the critical points where a sharp increase in cost may occur are individual to the application and cannot easily be generalized. Questions to be asked by DSN stakeholders faced with this problem should focus on applications that might need little precision and/or item-level adjustment. In case of ejecting a part versus loading a part, the costly difference is that in ejecting, the location, geometry, and orientation of the part in the milling system is precisely defined and readily available from the milling operation itself. Loading, on the other hand, requires locating, orienting, and precisely placing the part in the system—e.g., grabbing the part from a loose collection in a bin. This requires a vision-based system and high felicity in terms of the robotic tool.

APPLICATIONS, CHALLENGES, AND POTENTIAL BENEFITS OF AI, ML, AND ROBOTICS IN DSN

In this section, we will briefly look into numbers and applications associated with AI, ML, and robotics in DSN, as well as emerging challenges and benefits.

AI, ML, and Robotics in DSN in Numbers

There are many numbers and predictions floating around when it comes to the impact AI, ML, and robotics will have on the future of DSN. These include predictions that advances in AI, ML, and robotics will replace 800 million jobs by 2030.³ The report estimates a potential to automate roughly 50 percent of all current work activities by adapting already demonstrated technologies. At the same time, this does not mean that these jobs are “lost,” but that the nature of work will transform with the maturity of these technologies. If history is any indication, the introduction of new technologies has not reduced the overall number of jobs, but initiated a shift. The 2019 World Manufacturing Forum report highlights the changing skills that are required to compete on the future job market driven by AI, ML, and robotics.⁴

When we look at the predictions for the impact of AI, ML, and robotics on DSN and the associated numbers we see that the expectations are substantial. Predictions include a reduction of scrap rates by 30 percent, of annual maintenance cost by 10 percent, of inspection cost by 25 percent, reduction in lost sales by 65 percent, of annual downtime by 20 percent, of supply chain forecasting errors by 50 percent, of supply chain administration associated costs by 25 to 40 percent, as well as possible inventory reduction of 20 to 50 percent based on advances in AI, ML, and robotics.⁵ Other predictions include a 25 percent improvement of inventory turnover⁶ and at least 50 percent improved prediction of performance degradation across multiple manufacturing scenarios.⁷

AI, ML, and Robotics Applications in DSN

The expectations within DSN for AI, ML, and robotics to make a transformational impact are huge. Every area of any DSN has at least some possible application where AI, ML, and robotics are currently being considered, and most likely already tested. However, there are some applications that lend themselves to be a more natural fit (see characteristics of suitable applications cases discussed before). In this section, we provide a list of selected areas where AI, ML, and robotics will have an impact and that might present a good start when your DSN is considering AI, ML, and robotics implementations. While DSNs are complex and cover a broad scope, we tried to provide a diverse list of application examples and areas to inspire innovation. Again, AI, ML, and robotics are strongly intertwined and not distinguishable black or white in most cases. Therefore, the following three collections are mainly to provide some form of structure to the reader.

First, we will look into AI and ML application in a more traditional supply chain setting:

•   Data-driven supply chain planning

•   Route optimization and delivery planning using real-time data (e.g., traffic information)

•   Optimizing warehouse management (e.g., machine learning based forecasting of optimal stock level)

•   Decentralized control of packages and other items that plan their own route, adapting to the environment

•   Data-driven scheduling and prioritization of orders in distributed DSN

The following AI and ML applications are growing within the extended DSN domain:

•   Automated procurement algorithms that crawl the market and negotiate automatically

•   Data-driven demand forecasting (e.g., machine learning powered prediction of fashion trends based on unstructured social media data using natural language processing)

•   Automated risk monitoring and/or supplier qualification, monitoring, and selection

•   Increased product quality during production using data-driven optimization

•   Zero downtime based on data-driven predictive or preventive maintenance algorithms

•   Real-time tracking of orders by customers and other DSN stakeholders

Last we cover some typical physical automation applications within DSN:

•   Autonomous trucking (driving) for distribution

•   More inclusive workplaces with cobots supporting workers (including but not limited to workers with disabilities and an aging workforce)

•   Competitive manufacturing operations located close to end customers enabled by lower labor cost using robotics on the shop floor

Challenges for AI, ML, and Robotics in DSN

AI, ML, and robotics have come a long way and are already providing value to leading companies around the globe daily. However, there are several challenges that companies need to be aware of when contemplating the introduction of these technologies in their DSNs.

One major factor is data itself. On the one hand, consider the previously discussed “Vs” (see Chapter 3), but specifically data quality and data sharing. Data quality has a tremendous impact on the potential value of AI, ML, and also robotics applications within a DSN. Even when only a fraction of the available data is of low quality, that can impact the prediction accuracy and in the end the overall usefulness of the AI/ML solution. The other key aspect, data sharing, is essential within a DSN where several stakeholders are working together, both on the technical and business sides.

A more AI- and ML-specific issue is the often black box nature of many AI/ML predictions. Similar to manufacturing, in a DSN the stakeholders aspire to understand why certain decisions are made. Often a simple “the algorithm said so” is not good enough—most certainly not in cases resulting in serious injuries or financial repercussions. This problem is heavily researched, among others by DARPA’s XAI (explainable AI) initiative and several efforts to merge physics-based modeling with data-driven approaches. This is crucial for the social acceptance of AI, ML, and robotics and for trust to be placed in their outputs.

Still being comparably novel technologies, there are many technical issues that need to be addressed for them to become broadly adopted and accepted. These include but are not limited to distributed computing, hosting of services and data, real-time analysis in the cloud, latency, connectivity in remote areas, and use in “dirty” (abrasive, metal heavy, etc.) spaces.

The last challenge that we want to present is liability of AI, ML, and robotics in DSN. Partly connected to the black box nature of many of the algorithms, the question is still unsolved who is responsible if something goes astray. What happens when the forecasting algorithm’s prediction is wrong and the warehouse system runs out of stock, halting production? Or when an autonomous truck has an accident and harms humans based on a decision of the AI algorithm? These are questions that cannot be overstated and that need to be addressed carefully, bringing technologists, policy makers, lawyers, and other stakeholders together.

Benefits of AI, ML, and Robotics in DSN

After covering the challenges that have yet to be overcome, we will now briefly discuss the benefits AI, ML, and robotics offer for modern DSN—and there are plenty! There are many obvious ones, mainly around reducing the efforts and resources needed to do certain things (automating). Many of the applications mentioned earlier in this chapter can be grounded in such initiatives. These are more obvious, and we will not focus much on them but rather on the more forward-thinking and not yet fully developed ones to inspire the creativity of our readers.

Globally, the many developed and developing economies are faced with an aging population. The baby boomers are close to retirement, and with that, companies are struggling to find qualified workers with the required skill set. At the same time, many older workers would like to continue to be productive members of society—and AI, ML, and robotics can be an enabler to facilitate this. Cobots and AI-based augmented systems can provide the tools for older workers to safely and productively continue working in an industrial setting. Both features, safety and productivity, are key dimensions where AI, ML, and robotics are impactful and value adding to various areas within a DSN.

On a global level, the pressure to reduce the energy footprint and a more mindful use of resources while providing highly efficient industrial production to a growing population is a challenge that AI, ML, and robotics can be integral in addressing. It is almost impossible to predict and optimize the energy footprint of a global DSN without AI/ML, given the complexity and dynamics involved. In this case, AI and ML are the only viable option to address this pressing challenge of our planet.

SUMMARY

Summarizing, there are several exciting benefits associated with AI, ML, and robotics in a DSN environment. The technologies are mature and continue to develop at a rapid pace. However, adopting AI, ML, and robotics does not come easy. Therefore, we have put together 5 + 1 rules that interested decision makers in DSN need to respect (see Figure 4.12).

FIGURE 4.12   5 + 1 Rules for Successful Use of AI in DSN

^* The authors would like to thank Dr. Juergen Lenz for his support and insights in developing this chapter.