This is the second article in a series on Demystifying Machine Learning. If you have not read the first article yet, you can check it out here: Demystifying Machine Learning.

Reviewing Decision vs. Insight in Machine Learning

In our first article, we outlined the three main types of machine learning (supervised, unsupervised, and reinforcement learning) and identified what outputs could be expected from each. If you seek assistance or automation in making a decision that has a correct or incorrect answer, use a supervised machine learning model to aid in selecting the correct answer to new problems based on a training model using similar, past problems. If instead, you are searching for unknown insights in the existing data/situation then you should turn to an unsupervised machine learning model to help you discover previously un-examined patterns and anomalies.

Machine Learning Applications for Logistics Operations

In recent years, machine learning (ML) and artificial intelligence (AI) have been applied to a varied number of problems and industries ranging from health to entertainment to banking. One of the fields that stands to benefit significantly from the application of machine learning is logistics operations.

For large commercial organizations, governments, and militaries, the movement of people, products, and equipment is complex. Assuring items arrive in the correct location, at the correct time, in the correct quantities, is challenging even at a small scale. Some organizations, such as shipping and delivery companies, have utilized emerging technologies to conquer the complexities of these activities and can reliably and consistently manage a massive network of products and transportation equipment. Other organizations have had less success in trying to overcome these challenges.

Machine learning and AI can be used to predict transportation times, alert of equipment breakdown risk, or quantify the impact of changing weather on resource consumption rates, to name a few examples. This information can be the key to successful logistics operations activities. In the next sections, we will review some logistics scenarios where decision or insight support is achieved using ML techniques to improve the outcomes of logistics operations. These scenarios will focus on consumables management.

Logistics Use Case 1: Consumption Rate Prediction

Consumables, no matter their class (e.g., food, water, fuel, or ammunition), are a major factor in the success and complexity of the planning, execution, and sustainment of a mission. While an excess of consumables can complexify planning, transportation, and storage, a shortage can derail execution and jeopardize a mission and the engaged personnel. Finding the right balance between what is needed and what is manageable is key.

Supervised learning techniques applied to historical data from missions of a similar nature (e.g., same geographical operational area, weather conditions, objectives, etc.), can help decision makers to better predict future consumption and plan accordingly. For instance, historical water consumption data can be analyzed to identify trends and generate predictive formulas. These formulas can be applied to forecast data by extrapolating future consumption and support leaders in making educated decisions about resource requirements and sustainment operations when planning for future events.

Logistics Use Case 2: Uncovering Insights

While predicting consumption rates is a direct way of optimizing consumables management, uncovering insights about those rates is just as important. This can be achieved using supervised or unsupervised learning methods, both methods providing different types of insight.

On one hand, supervised learning can be used to analyze historical data to help understand the correlation between different independent variables (e.g., environmental factors) and an associated dependent variable (e.g., water consumption rates). For instance, when looking at past water consumption, this type of analysis can be used to identify and rank environmental factors (e.g., temperature, humidity, elevation) by the size of their impact on the resulting water consumption (this measure of impact size on the dependent variable is known as the correlation coefficient).

Although, it is often the case that two independent variables may be strongly related (known as correlation), . Two variables can be strongly correlated to one another, leading to an incorrect belief of causation between the two variables, when the correlation is actually the result of a third, unexamined, variable which is impacting the two examined variables in a process known as confounding. For instance, while ice cream consumption and the number of shark attacks are strongly correlated (i.e., they tend to follow each other), they are both caused by rising temperatures (the confounding variable):

  • high temperatures in the summer cause people to eat more ice cream,
  • those same high temperatures cause people to swim more, exposing them to shark attacks.

On the other hand, unsupervised learning can analyze datasets to identify clusters of similarities (i.e., data records following similar patterns), to support complex environments in which the amount of data variables cannot be processed by the naked eye. In a simple scenario with two variables (e.g., temperature and water consumption), data records can be easily plotted on a 2D chart that can be interpreted by the naked eye. In more complex scenarios, with more numerous variables (e.g., temperature, water consumption, humidity, and elevation), humans cannot represent and interpret the data themselves, and rely on artificial intelligence to provide that analysis.

By identifying similarities, it also becomes possible to identify anomalies, in the form of isolated records. When it comes to consumables, these anomalies are significant findings that could represent over/under consumption of resources, and usually indicate a need for further investigation to understand and identify the root cause (e.g., leaky equipment, faulty sensors, degrading performance).

Machine Learning Challenges

These analytical tools (ML and AI) and their associated benefits come with challenges that need to be properly understood to take full advantage of them:

  • A right balance between “enough” and “too much” needs to be struck when it comes to data volume. A dataset on the smaller end can be misrepresentative of the field situation by missing out on independent variables or low-frequency events. Larger datasets can be polluted with irrelevant variables or outdated data, diluting relevant data, and overwhelming the learning algorithms with noise that leads to inaccurate or less accurate outcomes.
  • The ML outputs, be it quantitative values or clusters, can be difficult to translate into actionable information. As discussed in the correlation vs causation section, results must be contextualized before deriving informed courses of action. This contextualization requires subject matter experts to interpret the environment captured by the different data variables to properly identify and distinguish causations from correlations.

Conclusion

In this article we reviewed the difference between decision and insight solutions generated by machine learning models and the importance of knowing which approach is appropriate for your scenario. We examined how machine learning can be applied to problems in logistics and consumable management, and how it can provide valuable information to greatly improve the efficiency and success of logistics activities. Lastly, we examined some key challenges that must be understood when utilizing machine learning solutions. In our next and last article in this series, we will discuss deep learning and look at how multiple types of machine learning models can be combined to form neural networks.

Nexus and its partner Black & Rossi are actively implementing AI and ML within the JEDI-X platform to enhance decision support and automation. Our experts can help you learn more about ML, the opportunities within your organization, and how to overcome associated challenges. Contact us now to speak with one of our experts.