Generative AI has been heralded as the most profoundly impactful technological innovation since the iPhone. ChatGPT in particular has captured the world’s attention with its ability to convincingly generate whatever text the user desires. It has passed numerous standardized tests, from the Bar exam to the SATs, and its essay-writing prowess poses an existential threat to the integrity of education itself.
Other tools have shown impressive results in generating art, videos, code, and more. Not surprisingly, many predict that these generative AI tools will be widely disruptive — particularly for industries like media, marketing and legal that deal with text and images. What’s less obvious is how generative AI will impact supply chains.
The truth is that text and image generation is just the beginning of what generative AI can accomplish. It can also be used to generate solutions to optimization problems that abound within supply chains.
Generally speaking, any situation where you have a wide range of possible solutions, and you want to find the best solution, can be thought of as an optimization problem. A simple example is when Google Maps tries to find the fastest possible route to your destination. While this works reliably well, for more complex optimization problems, classical computers don’t have an efficient way to find the best solution and can only generate approximate solutions.
In contrast, a generative model could be trained on the best existing solutions to an optimization problem – for example those obtained from classical heuristics or MIP solvers – and learn what makes a good solution good. Much like how ChatGPT learns from existing text to generate new text, this generative model could then generate new solutions to the optimization problem. We call this approach Generator-Enhanced Optimization (GEO).
Potential supply chain use cases include finding more efficient shipping routes, optimizing the organization of warehouses to speed up order-picking, or selecting the best combination of suppliers, distributors and vendors. Given the complexity of most global supply chains, there is ample room for optimization — and cost savings as a result.
It sounds promising, but for years quantum computing has also been touted for its ability to solve optimization problems, and yet today there is not yet a documented example of quantum computers providing an advantage for optimization. However, generative AI may be the fastest avenue to realize that quantum advantage. It may also be the first place we see a practical quantum advantage at all.
Quantum-enhanced Generative AI
To vastly oversimplify things, generative models like those behind ChatGPT work by learning patterns in massive datasets and producing new data that conforms to these patterns. In other words, they learn to replicate the probability distributions of the training data. Quantum computers have the ability to encode and sample from complex probability distributions in a way that classical computers cannot, giving them a potential advantage in generative modeling.
How is this possible? For one, quantum entanglement can encode distant correlations within a dataset in ways that would be difficult for a classical computer to simulate. Secondly, the inherently probabilistic nature of measuring a quantum state makes quantum computers the ideal vehicle for sampling from probability distributions.
The end result is the ability to generate a more diverse range of solutions to the generative modeling task. In the context of optimization, this means quantum generative models could generate new, previously unconsidered solutions.
But there’s a catch. Quantum devices are currently limited by low qubit counts and high error rates. But we don’t necessarily need quantum devices. However, tensor networks, originally popularized among quantum physicists for simulating quantum states on classical computers, can be used for generative modeling today. And as quantum hardware matures, these quantum-inspired models can be translated into real quantum circuits, making them forward compatible with future quantum devices.
Tensor networks have shown particular value for optimization problems with equality constraints. An equality constraint is a condition that must be satisfied exactly for the solution to be valid. Without a way to natively encode these constraints, traditional optimizers can generate many invalid solutions, resulting in inefficient and expensive searches.
On the other hand, tensor networks can be constrained in a way that only outputs valid samples, resulting in the generation of more novel and high-quality solutions to optimization problems. And while equality constraints can worsen the performance of other quantum or quantum-inspired approaches, the opposite is true with constrained tensor networks, which deliver better computational performance at a cheaper cost for each additional equality constraint.
Optimizing the Supply Chain
There are many possible applications of GEO that could make the supply chain more efficient. Below are a few examples:
- Supply Chain Optimization: Selecting suppliers, distributors and vendors to maximize product quality and demand coverage while minimizing costs and delivery times.
- Delivery Routing Optimization: Identifying the most efficient delivery routes to reduce costs and shorten delivery times.
- Warehouse Optimization: Designing warehouses to increase efficiency in order-picking.
- Facility Location Optimization: Identifying the ideal locations for supply chain facilities to reduce construction and transportation costs.
- Workforce Scheduling Optimization: Balancing workloads for multi-commodity teams while minimizing labor costs.
- Reverse Logistics Optimization: Scheduling logistics for recycling, reuse, disposal and product recalls to minimize costs.
Of course, supply chains can vary widely from industry to industry. You may have additional optimization use cases that are unique to your business. But across the board, generating better optimization solutions has the potential to reduce costs and speed up the supply chain. Optimization could also reduce waste and cut carbon emissions — a great place to start for businesses looking for ways to reduce their carbon footprint.
How great is the potential value at stake? The only way to find out is to try. We are still in the early days of generative AI — and even more so with quantum-inspired generative AI. By building and deploying generative AI applications, not only do you stand to gain a competitive advantage, but you may also make discoveries that advance the field.
It’s important to reiterate that tensor networks are forward-compatible with real quantum computation. Businesses that deploy tensor networks for optimization may not only gain an advantage today, but they would also be in position to gain a potentially greater advantage as quantum hardware becomes more powerful. In other words, they will become quantum ready.