In 2018, Google’s AI team entered AlphaFold into an international contest called the Critical Assessment of Structure Prediction. This contest, which had already been running for 13 years, was designed to inspire a major innovation in medical science. While it might not sound very exciting, identifying protein structures is a time consuming and expensive process that chokes the development of new medicines. AlphaFold, however, took a task that typically involved decades of work and completed it in just three months. It did so using a principle called deep reinforcement learning (deep RL), a machine learning process built to mimic how humans think and act – through unassisted exposure to stimuli. Though we’re living through a peak of generative AI experimentation, that’s not the best place to look for groundbreaking innovations in manufacturing. And the reason why is more straightforward than you’d think.
If we take the deep reinforcement learning techniques utilized by models like AlphaFold and apply it to the manufacturing industry, there are enormous opportunities for innovation. This sector, with its complex challenges in production optimization, product development, and supply chain management, is ripe for the transformative power of AI. Deep reinforcement learning can revolutionize manufacturing by finding novel solutions to age-old problems—enhancing efficiency, reducing waste, and even pioneering new materials and techniques. This approach offers a significant shift towards leveraging AI not just for incremental improvements but for industry defining advancements that break open what’s possible in manufacturing, signaling a new era of industrial innovation.
Deep Reinforcement Learning Focuses On Progress Toward Targets
At its core, generative AI tools like ChatGPT and Bard aren’t built to understand progress. In simple terms, they work through a very complicated system of pattern recognition: after seeing millions of Linkedin posts or song lyrics, the system begins to recognize the core features. When you ask it to write a LinkedIn post, it makes a prediction based on all of the examples it’s been given beforehand. There is performance evaluation, but the core system of many AI tools like ChatGPT are fine tuned with feedback while maintaining their initial training rules. Deep reinforcement learning, on the other hand, offers more flexibility and is built with an end goal in mind and no rules on how to get there. Deep RL utilizes ‘deep learning’ – neural networks that mimic how humans connect experiences together to form conclusions – and ‘reinforcement learning’ – a type of trial and error performed on the scale of thousands at a time. Deep RL models simply start working and don’t stop until they’ve learned the best way to reach their goals based on their own experience. Creating a model that could identify the best locations for emergency stop systems means that it will test every possible placement and orientation to create the best outcome.
The easiest way to understand this is with video games. Imagine you set up a deep reinforcement learning model on the original Super Mario Bros. Through millions of simulations, the algorithm learns from each interaction—understanding the repercussions of hitting a block or falling into a pit. This iterative learning process, driven by trial and error, enables the model to master the game and achieve what seems like an impossible feat: navigating complex levels with precision. Over time, it would learn how to complete entire levels of Super Mario Bros without ever accessing existing data or advice from a human observer. Applied to material science or even factory planning, you can see how this kind of trial and error can yield powerful and unexpected results.
Deep Reinforcement Learning Creates Original Ideas
This is why deep reinforcement learning has so much potential for manufacturing innovations. Rather than predicting what something should be based on examples, deep reinforcement learning teaches itself to find a solution. In complicated situations with many paths to a desired goal, AI models can uncover solutions that aren’t a part of the conventional understanding. This approach not only enhances the efficiency and sustainability of manufacturing operations but also pushes the boundaries of what’s possible, enabling companies to achieve breakthroughs that redefine industry standards and consumer expectations. In essence, deep reinforcement learning acts as a catalyst for innovation, driving the future of manufacturing towards uncharted territories with raw creativity and unguided learning.
To be fair to generative AI, there is a lot of real value to be had with tools like ChatGPT and Bard. As Deloitte finds in their “Generative AI Dossier”, generative AI can create more efficient processes across HR, Marketing, and Customer Service. Such tools can easily personalize content to users, refresh and tweak existing work, and empower a writer to work faster. You might be able to find novel concepts in generative AI by combining two fields that don’t typically go together (“write a Muppets space opera about cream cheese”), but generative AI will always create things based on what it already knows. Even if you find truly valuable ideas, any competitor can retrace your steps to replicate your work with relative ease.
The Catch? Deep Reinforcement Learning takes a lot longer
Real innovation doesn’t happen on the first try. Or, with deep reinforcement learning, maybe not in the first 114 million tries. But in the thousands of simultaneous simulations it will try every possibility to find the best answer. Both the advantage of deep reinforcement learning and the biggest drawback is the methodical, drawn out process it uses to find the best answer. Provided that it was set up correctly, deep reinforcement learning will eventually uncover massive benefits for your manufacturing products and processes. This time-intensive and resource-heavy process can be daunting, but the potential for creating novel, industry-leading technologies and solutions far outweighs the initial investment.
Deep Reinforcement Learning Only Works with Pass/Fail
Because deep reinforcement learning is not given information at the start, it’s very hard for it to understand complex and nuanced goals. This allows the deep reinforcement learning model to continue to iterate on its successes and avoid failures without getting too caught up in juggling different evaluations of the same outcome, an important factor when incorporating millions of results. While this will limit the scope of what it can be applied to in manufacturing, there are still plenty of applications with specific and measurable goals that are ripe for innovation. Quality control inspections, assembly line optimization, or predictive maintenance are just a few examples of pass/fail processes in manufacturing. You could ask it for the optimal time to replace shock absorbers or the optimal shock absorber placements and energy capacity to maximize durability. The true range of deep reinforcement learning depends on your own ingenuity and creativity in how you deploy it.
Incorporating generative AI into current manufacturing processes can significantly enhance productivity and maintain a competitive edge. But if you’re interested in becoming a leader in the manufacturing industry and transcending traditional methods, deep reinforcement learning offers a unique and compelling opportunity. This strategic adoption could create a transformative leap towards mastering complex production challenges, optimizing supply chains, developing novel products, or even revolutionary manufacturing techniques. Investing in deep reinforcement learning partnership for your company places you at the forefront of technological adoption, shaping the future of manufacturing.