Making Machines Learn Like Humans in Planning Supply Chains
Machines can learn by imitation or observation like humans do. In other words, looking at many instances of occurrence of an event and then relating that occurrence to some other patterns of data exposed to the machine. For example, the ice cream sales are low in winter can be detected by its correlation to lower temperatures, increase in sugar prices, or the scarcity of milk. If enough data is given, then ML algorithms can predict the sales of ice cream every winter. Humans come to the same conclusion given the same data. However, they are capable of reasoning at a deeper level by connecting to their existing models (understanding) of the world that may not be available to the computer. In the case of a drop in ice cream sales, they can conclude that there will be lower revenue for the industry. Then they evaluate this outcome to see if it is desirable or not. Next, they fall back on their current experience (what other things they have learned), and search for ways of improving revenue. Their experience and knowing can make a difference in what solutions they can come up with. Some may say it is what it is. Some look deeper and say if it is the weather, then lets promote sales in regions where the weather is warmer. Or if it is due to sugar prices, then look for alternative sweeteners. Can a computer do the same? Yes, if they have the data made available to them. Can the computer come up with an advertising campaign that promotes ice cream to healthy living in winter? Not by itself unless it knows advertising has the property of increasing revenue.
In the above simple example, humans have learned about a solution not just the cause. This is the real intelligence that can also be given to machines using LLM by understanding the properties or attributes of objects. In supply chain planning, the ML algorithms warn about issues such as high or low forecasts, maintenance occurrences, or even inventory levels. However most, if not all, cannot come up with solutions that can be prescribed to the end users. To be able to prescribe, computers need a full understanding of the world in terms of shape (structure or digital twin) and properties of the objects, what we call behavior. The shape of a table is understood, however its properties are many, including to bear weight, to protect against earthquake, and to have a thanksgiving dinner.
Consider a supplier which is always late at certain times of the year depending on the quantity they are supposed to deliver; or perhaps their price increases on a regular basis compared to other vendors. Systems need to understand the objectives of the organization so that when they learn about this supplier then they can act according to the objectives. A true digital twin may give an alternative supplier with a lower cost to be used instead. However, the alternative may have longer lead-times and/or lower quality or even higher carbon emission standard. Humans fall back on their values and objective to decide what action to take. Systems need to understand such objectives and values to decide. This goes beyond having a digital representation of the world in a digital twin. This requires a digital twin of the behavior of the physical world not just the appearance, or structure, of it.
To this end, we believe both structure and behavior are needed to have a truly intelligent system to quickly provide answers and recommendations to the management. The use of attributes and Attribute Based Planning provides a means by which this behavior can be represented and learned respectively, given the objectives of the management and the organization. The objectives are of course profit, market share, carbon emission reduction, cost reduction among many others. Our approach has been to provide the flexibility of stating multiple objectives in a way that the system satisfies an overlapping area of all the objectives, considering the tolerance of each objective. Hence, creating an environment for negotiation amongst preferences. This is not a majority wins kind of situation but finding a solution that meets as many of the required objectives as possible as they are constantly changing, as does the behavior of the organization.
For real intelligence both structure and behavior are needed to be represented and learned, in real world, humans are constantly changing and learning and improving. For systems to be able to do the same, they need to have the capability to have the same capability to change both as the real-world changes. For more information on the use of AI and ML as well as behavioral representation of supply chains using Attribute Based Planning click Here.
