Published in


AI in Logistics (2/2): Challenges and Trends

A data scientist’s view on logistics in the post-COVID era.

A trailer truck by cloudy weather. Source: Robson Hatsukami Morgan.

Beyond the large range of use-cases that we have tried to list, it is equally important to understand what are the fundamental challenges that logistics have to deal with. Again, we put aside robotics and autonomous vehicles as there are no doubts that they will keep playing an increasingly important role and that this trend is already clearly established.

COVID-19 and the Non-Stationarity of Time-Dependent Data

Our economy is an open dynamic system and logistics is a substantial subset of it. This means that things can change, sometimes abruptly, especially when exposed to a massive exogenous shock like the one induced by the lockdowns decided during the COVID-19 pandemic.

One can try to model the response to this shock to some extent. However, that still leaves us with a rather practical problem: how to generally deal with the non-stationarity of processes when using machine learning models for predictions? This kind of shock induces strong non-stationarity, when the statistical distributions of considered quantities suddenly change, which can make some machine learning models instantly obsolete.

Non-stationary time-series can drift indefinitely. Source: Wikipedia.

There can be ways to mitigate this problem. For example, by:

  • Reducing the prediction horizon, a way to admit one cannot confidently make predictions,
  • Integrating data that contain such shock and more context whether these data are historical or simulated (back-testing combined hedging against tail risk),
  • Having service architectures that effectively retrain machine learning models on a regular basis to cope with new conditions.

This is routinely done in the financial world. But the problem is broader than that and calls some serious reflection on the level of resilience (or antifragility) one should expect when experiencing so-called Black Swan events and on the true nature of uncertainty and how to quantify it. For an advanced discussion on the topic, one can only recommend the work of the mathematician Benoit Mandelbrot or the popular book series by Nassim Nicholas Taleb.

Globalized Optimality versus Regional Resilience

In the last few decades, the trend of global integration of supply chains has been to favor a certain optimality, which is ultimately the minimization of costs in the short or middle term. However, these short-term benefits increased the fragility of those supply chains whose resilience has been largely neglected. The question then is how to price in the long term benefits of a more resilient supply chain? How logistics operators should balance the optimality of their operations versus the robustness of their business?

These are very complex issues and the current trend seems to be the reintroduction of regionalization in a truly multipolar world. In other words, we seem to witness a large wave of decentralization. But technically speaking, balancing costs optimality and resilience leaves us with some advanced multi-objective optimization problems which are by nature ambiguous (but not intractable) and can be extremely difficult to formulate in mathematical terms due to the intrinsic complexity of the domain of logistics itself. Hopefully, the progress being made in AI, especially in the field of reinforcement learning, will help to further tackle these issues more efficiently.

The rise of AutoML and Reinforcement Learning

Historically, the usage of advanced heuristic algorithms has been very successful in solving problems typically encountered in logistics such as the traveling salesman or the bin-packing problems. When properly used, heuristic algorithms can be extremely fast and provide near-optimal solutions to relatively advanced optimization problems.

Yet, there are still some shortcomings: the solutions provided by such heuristic algorithms can be counter-intuitive (a quite common UX problem) or when a machine-learning formulation is more adequate but hard to implement because of complexity or lack of data or when the optimization problem itself is too hard to be explicitly formulated (e.g. advanced warehouse inventory optimization). These are the realities of applying AI to real-life problems.

When implementing predictive models for time series forecasting which can demand a lot of work from ML engineers, the AutoML framework could be very helpful as it enables to find the best models in a nearly automated fashion. It is especially useful in the context of non-stationarity of data as aforementioned or when the number of models to be trained is very large. The downside is the explainability of the final models. Explainability of AI is still subject of a large amount of academic research activities…

A very simplified view of the Reinforcement Learning paradigm. Source: Coach.

Finally, reinforcement learning seems to be the most promising for logistics as it is the only approach that can efficiently solve highly complex optimization problems like organizing a supply chain or improving the optimality of a warehouse inventory. Moreover, it may offer more interpretable solutions than other approaches. But there is no magic here either: reinforcement learning can be extremely demanding in terms of computational resources as the system needs to try out many possibilities and it necessitates a very realistic simulation of the environment’s response to certain conditions which is far from trivial. Yet, some remarkable progress is to be expected here.

Towards integration with blockchain-based tracking and settlement platforms?

Complex logistics operations and large supply chains demand trust and transparency which require lots of effort to implement. The speed of settlements is also a problem as payments can typically take weeks or even months to be processed. Yet, we saw that AI can improve the efficiency of operations tracking by large margins while speeding up and reducing human errors in back-office tasks.

A parallel quickly growing technology are the blockchain applications dedicated to business and supply-chain management like VeChain. Large players are already involved in developing these so there is little doubt that blockchain applications for logistics will be part of the daily landscape soon enough. They will enable an increase in the level of trust in a more or less decentralized fashion while making all kinds of settlement transactions near-instantaneous. This means that the AI systems involved in tracking operations will have to be integrated into blockchain oracle systems that can trigger checkpoint validations, settlements and other follow-up steps automatically or with little intervention of humans.

The implementation of these in real-life applications still remains to be seen but all the necessary components already exist.

Building User-Friendly Solutions and Trustworthy AI

Finally, perhaps less visionary-sounding but not less crucial are the topics of making user-centered and more broadly human-centered applications. The questions are:

  • How do we know that the AI-based solution implemented is the most relevant?
  • How to ensure that the final users get empowered by this application and actually use it?
  • How to prove that an AI application is technically reliable, regulatory compliant and generally truly serves the purpose it was intended for?

These are difficult questions that have to be addressed in the most thoughtful and systematic way.

The usability of AI solutions can be well-covered by the usual principles of UX and Service Design and will be the subject of a later article. In practical terms, it means that designers closely working with data scientists and machine learning engineers should be able to acquire the necessary understanding of what “Design for AI” takes.

As for the trustworthiness of AI solutions, it is a broad topic with deeper technical implications but the most worthy approach is the one that will allow us to comply with the future round of EU regulations. For that purpose, the ALTAI document published by the European Commission provides useful guidelines for stakeholders.


2020 was a shock for logistics and supply-chain management, triggering tectonic shifts in the ways to approach the organization and planning in these fields. These changes will be largely fueled by advanced AI solutions.

Logistics will be more regionalized, decentralized and hence much more resilient thanks to the higher speed, adaptability, transparency and trustworthiness of integrated AI solutions.

Looking for the first part of this series?
For a discussion of use-cases for AI in logistics, have a look at part one:
AI in Logistics (1/2): Use-cases.

This article was written for Sclable’s blog on Medium.
If you liked it, give it a clap and share if you ❤️




Combining Design, Data and Code. From strategy to scale. Based in Vienna, Austria. Say hello at or

Recommended from Medium

Human-Centered AI: Reliable, Safe & Trustworthy: HCIL Symposium

The Next Touch Point: AI and Culture-IQ

Introducing Constella Intelligence

What Error Correcting Codes, Artificial Intelligence, Going to Sleep, and Free Speech Have in…

How Do Robots Learn to Walk — Localization Using Particle Filters

How the Future of AI Will Be Owned by All

AI and the Digitization Journey of Contracts

Modeling the way towards Artificial General Intelligence

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Charles Dietz

Charles Dietz

Data Scientist & Physicist

More from Medium

Holy Grail for ICE engines ! Homogeneous Charge Compression Ignition (HCCI) Engine

Are AI And IoT Development Services Changing The Future?

AI in Biotech: A Simplified Overview

Rendition of DNA molecule

A Look Inside the World’s First Law Regulating Autonomous Driving (Bonus)