AI in the Supply Chain Management: Where are We Now?

The 'travelling salesman problem' (TSP) hails back to the early 19th century. It has since become a minor obsession in computer science and, more recently, in machine learning research⁴ aimed at improved route design.

The theoretical challenge is to find the optimum route for a series of stops over a wide geographical area, conserving time and resources where possible. The best solution is surprisingly elusive, and is susceptible to a wide range of mathematical approaches:

Although routes can be optimized on a case-by-case basis, there is an existing general algorithm that's able to predict the best possible way from point A to point B.

Developed in 1976, the 'universal' algorithm's best route is still 50% longer than the most efficient route for any specific case⁵. It took the scientific community more than 35 years⁶ to find ways to improve this generalized algorithm even by 10%.

Over the last decade, however, the combination of big data and GPU-accelerated machine learning has made new inroads on route optimization.

UPS's On-Road Integrated Optimization and Navigation (ORION) framework currently spares the logistics company 100 million miles of travel per year over 55,000 routes in the US, saving ten million gallons of consumed fuel and 100,000 metric tons of carbon emissions⁷. The system now constantly analyzes vehicle feedback in a centralized machine learning framework⁸ based in New Jersey.

Other AI-based initiatives for optimizing supply chain routing include:

• Amazon Prime, which leverages Amazon research into artificial intelligence, machine learning, and predictive statistics to calculate total deliveries and allocate vehicle space based on weight, volume, driver workload, traffic, and even the type of building that deliveries are destined for⁹.
• LaMP, an AI-powered regional logistics platform covering Southeast Asia that uses machine learning to consolidate deliveries from multiple companies into the most convenient vehicle available to deliver all of the items¹⁰.
• Uber Eats, which addresses the logistical complexity of hot food delivery by feeding GPS and motion data into a machine learning system that's capable of calculating the many variables that make efficient takeout logistics harder than transporting passengers¹¹. The system uses conditional random field (CRF) to model the 'trip state' and generate predictions for delivery times, as well as to calculate optimal routes.

In Southeast Asia, the growing demand for deliveries in densely-populated cities is leading to further innovation from the scientific community. Research out of the Singapore Institute of Manufacturing Technology¹² proposes a new method to overcome the logistics issues that arise when suppliers have incomplete information about peak demand times, the actual volume of customer packages (rather than the registered weights), and whether or not their delivery requirements will need to be redirected to third-party companies.

A study from Thailand uses artificial neural networks (ANNs) to generate a new guidance system for automated vehicles and robots that must traverse vast factory floors and navigate novel obstacles that may or may not be permanent.

Finally, a collaboration between industry and academia in China recently published details of a smart tour route-planning algorithm for popular Chinese cities¹³. The system mines historical data and is powered by a Naïve Bayes classifier. It can develop unique routes based on the interests of particular assemblies of tourists (rather than identifying and touring the 'most popular' locations).

Maritime transport represents around 90% of global trade¹⁴, as it remains the most economical method of delivering goods. Since it also represents 2.5% of global greenhouse gas emissions¹⁵ and was estimated in 2017 to lose US$18 billion annually to 'wait times'¹⁶, the shipping sector is motivated to achieve more optimized ocean journeys and make shorter and more productive stops at busy ports.

It's possible to incorporate historical data of the behavior of commercial ships near busy ports¹⁷ in order to assess the risks of being delayed there in a variety of possible circumstances. One marine information services provider uses¹⁸ the Random Forest Classifier in scikit-learn on such data to analyze historical patterns of 'loitering' and 'anchoring' near busy ports, in order to determine optimal routes and schedules.

Late-arriving ships have an impact on the tight schedules of ports and can compromise delivery targets. A proposal from Korea¹⁹ uses historical AIS transponder data to develop a more accurate model of the water-resistance for generalized shipping traffic — a critical factor in determining a vessel's actual running speed and estimating a more precise arrival time.

The system, designed to cut fuel consumption, correlates AIS data with live weather feeds to calculate the vessel's speed over the ground (SOG) and uses a variety of machine learning techniques, including gradient boosting regressors (GBRs), linear regression (LR), polynomial regression, extra trees regressors (ETRs), and random forest regressors (RFRs).

Machine learning can help with the unloading and deployment of a vessel by improving scheduling, internal routing, and general readiness for the port's particular technical environment:

• Research funded by the Ministry of Science in Korea²⁰ proposes a novel system to speed Interterminal Truck Routing (ITT) optimization with the use of deep reinforcement learning via a Deep Q Network. The research analyzed 576,000 combinations of states and actions in order to develop an AI-driven model of the Busan New Port terminal in South Korea.
• The Port Of Montreal has adopted machine learning to optimize freight planning by comparing the ways that a number of loose variables (such as storage capacity, staff availability, vessel arrival times and rail deliveries) impact the fluidity of port activities²¹.
• German port operator Hamburger Hafen und Logistik (HHLA) uses machine learning to improve container location selection and stack location, to reduce 'dwell time' for unloading vessels at a major port²². The initial database model was generated from historic data from container handling operations and is now constantly updated. The developers foresee a long-term improvement in prediction accuracy of 33% for outgoing traffic and 26% for predictions of dwell time²³.

While a recent industry report estimates that only 12% of logistics businesses have currently adopted AI technologies³⁰, the global market of warehouse robotics is forecast to grow at a CAGR of 27% from US$ 6.12 billion in 2019 to US$25.8 billion by 2025³¹.

A key area for AI in warehouse automation is Automated Storage and Retrieval Systems (ASRS), wherein machines deposit and retrieve items from inventory.

Emerging machine learning approaches are leveraging cutting-edge sensor technologies including computer vision, haptic feedback, 3D vision, 'Light coding' with lasers, and laser triangulation. Additionally, ASRS is perceived as beneficial to wider fields of machine learning and to the development of AI benchmarking standards³².

After nearly a decade of research into warehouse logistics, Amazon is now deploying increasing levels of AI, including the automatic tracking and scanning of items³³, into its fulfilment centers. Though the company has stated that fully automated warehouses are a decade away³⁴, it has made waves with the introduction of the Italian-made Carton 1000 automated packing system at a number of locations³⁵ and has a long-term commitment to increasing AI-based deployment logistics automation solutions³⁶.

Pallet detection and tracking is another popular area of research in this sector. With over five billion wooden pallets in circulation in the EU and US combined³⁷, computer vision is able to identify individual pallets from surface wear and wood grain and even to assess when they have reached end-of-life³⁸.

While representing a fair bit less than 20% of actual deployed and functional AI-based SCM technologies, self-driving/piloting vehicles arguably occupy more than 80% of popular headlines around AI in the supply chain. For example:

• In 2016 industry participants in the European Truck Platooning Challenge completed a cross-continental journey with a convoy of robotic trucks shepherded by a human driver in the lead⁴³.
• One US self-driving truck startup has just launched a 4D Lidar system on its fleet of self-driving trucks (all of which must have a human operator) after a $350 million funding round in 2020 (see image below)⁴⁴. Level 4 (driver-free) autonomy is anticipated in the near future in restricted environments and weather conditions.
• Rolls Royce's autonomous shipping program⁴⁵ is the current emblem of a growing market⁴⁶, with its autonomous marine systems subsequently absorbed into Kongsberg Gruppen's ambitious plans for unmanned short-range cargo vessels⁴⁷.
• IBM's 'AI captain' took control of a crewless research vessel off the coast of the UK in 2020⁴⁸.
• Airbus reports the successful development of unmanned aviation, though customer confidence in a pilot-free airplane is currently lacking⁴⁹, while legal oversight and insurance costs are further obstacles⁵⁰.

If we don’t dwell on the topic here, it's because the pace of development in this sector is far, far ahead of the AI regulation environment and the revised definitions of 'liability' necessary for unsupervised, long-distance SDVs to embed deeply into the supply chain; and because the necessary future developments around this sector are at least as sociological and political as they are technological.