In the first installment of our three-part blog series on Explainable Artificial Intelligence (AI), I delved into the promise of AI in business processes—particularly in supply chain management—and how to build trust in the technology. I also took a trip backward in time to discuss the history of decision-making. But how dependable are today’s AI modeling techniques? Are we inadvertently overlooking their inherent limitations? In this second installment, I’ll explore the constraints of our traditional AI modeling techniques, highlighting areas in which they might fall short and further discuss the promise of explainable AI.
The models that have long been used for predictive modeling of complex systems, such as supply chains, have, by necessity, been simplifications of the true behavior. These models and algorithms, such as MRP II, have made it possible to generate plans across complex systems—and to a high degree of detail with a direct impact on many metrics (better customer service, lower inventory, higher capacity utilization). But, we have a misplaced belief in both the fidelity of the current models and the accuracy of the results.
This leads to situations in which new technology is required to replicate the results of the existing technology, as there is a misplaced belief that the existing technology represents the supply chain accurately and that the solvers calculate an optimized supply plan. Neither is correct.
The simplifications used to model the supply chains and calculate plans have come in three forms:
To reduce the complexity of the models required to represent the supply chains, the models are aggregated in the dimensions of time, geography and product hierarchy.
The problem with aggregation is that we often do not understand the consequences of the simplifications we perform. Just because products can be lumped into certain categories from a market perspective does not mean that their demand or supply behaviors can be grouped. Often, the demand and supply behavioral differences within a product group or family are greater than the differences between product groups.
The dominant use of mechanistic models in manufacturing and supply chain management is a significant limitation, as it ignores demand and supply variability. Traditional supply plans assume a single expected value for demand and supplier lead time, failing to account for real-world fluctuations. Upside, downside and commit scenarios are often used, but they are typically educated guesses, not data-driven responses to variability. Safety stock is the primary method to address variability, but it’s often applied at a single supply chain level and doesn’t consider strategic inventory placement. Excess capacity is an effective way to manage variability but is rarely integrated into the overall planning process. Bridging these gaps is crucial for more resilient supply chain management.
The majority of supply plan optimization tools rely on linear programming, even though many aspects of supply chains exhibit non-linear behavior. An approximation technique, piece-wise linearization, is often employed to address this limitation. However, there’s a trade-off involved: using more “pieces” to model non-linear behavior results in longer solving times, while using fewer “pieces” leads to less accurate models and consequently less reliable results. Although non-linear solving techniques can provide more accuracy, they tend to prolong solution times, though recent years have seen significant progress in this area.
Traditional modeling approaches such as aggregate modeling, deterministic modeling and the linear assumption have long been favored for their simplicity in explaining model outcomes. These methods offer a clear and interpretable framework, making it easier to comprehend the cause-and-effect relationships within the model. However, as innovative developments in modeling continue to push boundaries, they often challenge these constraints. While these advancements unlock the potential for greater accuracy and realism in capturing complex real-world phenomena, they come at a cost. The intricacies introduced by these innovative techniques can render model results significantly more difficult to understand. This complexity arises from the incorporation of non-linear relationships, interdependencies and intricate data-driven algorithms, making it increasingly challenging for human intuition to grasp the underlying dynamics of the model. Thus, as we venture into the realm of cutting-edge modeling, we gain accuracy but often sacrifice the simplicity and transparency that traditional approaches offer.
The key challenge lies in the fact that numerous AI applications use machine learning (ML) functions, rendering minimal to no evident understanding of the pathways they traverse to arrive at their conclusions. For scenarios involving relatively harmless yet frequent decision-making, like an online retail recommender system, an inscrutable yet precise algorithm emerges as the commercially sound strategy. This paradigm resonates extensively across the contemporary landscape of enterprise AI, wherein the paramount objective involves presenting tailored advertisements, products, social media content and search outcomes to the appropriate audience at the optimal moment. The rationale behind the “why” remains inconsequential, as long as it aligns with revenue optimization.
This paradigm has propelled an approach in which accuracy takes precedence above all else, emerging as the primary goal in machine learning applications. The primary users and researchers—often situated in distinct sectors of the same expansive technological entities—have directed their focus toward cultivating progressively potent models.
While we have grown accustomed to algorithmic decisions shaping our experiences—even in the face of possible biases—the deployment of AI in risk assessments within a business context, pivotal diagnostic judgments in healthcare, or the optimization of your supply chain policy has cast a glaring spotlight on this concern. Given the substantial implications in play, AI systems responsible for decision-making must possess the capability to explain their rationale.
Explainable AI represents an implementation of machine learning that offers a level of interpretability sufficient to provide humans with a qualitative and functional comprehension, often referred to as “human-style interpretations.” This comprehension can assume a global scope, granting users an understanding of how input features (referred to as “variables” within the machine learning community) impact the model’s outputs across the entire spectrum of training examples. Alternatively, it can adopt a local perspective, offering insights into specific decisions.
The premise of Explainable AI revolves around clarifying the reasoning behind decisions. This enhances the interpretability of AI models for human users, enabling them to fathom the rationale behind a particular decision or action taken by the system. Explainable AI contributes to the establishment of transparency in AI operations, potentially enabling the unveiling of the complete decision-making process, previously concealed within a black box. This revelation is presented in a manner that is readily understandable to human beings.
Diverse groups possess varying viewpoints and requisites concerning the degree of interpretability desired from AI systems. Executives have the responsibility of determining the essential assurances required to establish a sound decision process. They demand a suitable ‘shield’ against unintended consequences and damage to reputation. Management necessitates interpretability to foster comfort and cultivate confidence in system deployment. Developers, in turn, require AI systems to be explicable to secure approval for transitioning to production. Users, including both staff and consumers, seek assurance that AI systems accurately execute or influence appropriate decisions. On a broader scale, society at large seeks assurance that these systems operate in alignment with fundamental ethical principles, particularly in areas like mitigating manipulation and bias.
While the traditional models have served us well, we are standing on the cusp of a transformative phase in predictive modeling. It’s essential to acknowledge the shortcomings of our time-tested models and embrace the promising horizons offered by AI and its subsets. Explainable AI is not just about making machine learning more interpretable; it’s about building trust, ensuring transparency and allowing for ethical oversight in systems that are becoming increasingly integral to our daily lives. As we integrate AI deeper into our operations, it becomes imperative to prioritize understandability alongside accuracy. The future of modeling is not just about predicting outcomes but understanding the how and why behind them.
But how explainable should AI be? Find out in the next blog post, the third and last of our three-part series on explainable AI. The final installment will also explore exactly how to make your model explainable, the techniques that can be used to explain a model and, crucially, the role of a digital supply chain twin like Bluecrux’s Axon in bringing AI to your organization.
