INNOVATIONS IN SIMULATION
The Convergence of Simulation and AI in Manufacturing
If a person standing on a street corner opens an umbrella, could they cause a car accident?
Ask an average person and the answer is almost certainly ‘no’. Ask the same of an average AI algorithm, though, and you might be surprised to find that the answer is ‘yes’.
Why is it so?
An AI would be capable of connecting the number of people who are opening umbrellas with the number of car accidents. It wouldn’t take more than a couple of milliseconds to identify that the number of car accidents increases substantially once people start opening umbrellas – the correlation is clear and undeniable.
An average human, on the other hand, understands that this strong correlation between umbrellas and car accidents is caused by a third element: the rain. When it rains people open umbrellas, and when it rains, they are more likely to crash their car.
Does this mean that the human brain is always superior to AI or modern machine learning (ML)?
Not at all.
AI can make forecasts without needing to understand causes and this can be incredibly valuable. However, there are also some drawbacks to relying on AI and ML correlations alone, and those drawbacks are the reason that modeling and simulation turn out to be so complementary.
AI in Manufacturing: Data and IoT, Machine Learning, and Simulation
AI has applications well beyond predicting car accidents, of course.
In the industrial world, innovative firms have long used AI in manufacturing to identify connections between different parts of a production line, to optimize workplace processes, and to understand demand sentiment to better plan the supply to meet it. The capacity to rapidly parse incredible amounts of data and to identify patterns in that data is beyond the ability of any human and, in complement to human decision makers, has been the source of significant advances in operational practice.
Yet there are limits to even the most developed AI and ML algorithms.
Both AI and ML approaches in manufacturing rely on huge pools of data in order to make predictions about the future based on trends, patterns, and correlations in that data. Increasingly, that data comes from IoT devices that track, measure, and report to businesses seeking a return on that data.
Yet these predictions always come with a significant asterisk: not only are they ultimately based on umbrella/car accident-like correlations, but they have no capacity to predict anything that cannot be found in the data.
Put another way, ML has real difficulties in predicting anything that has never happened before.
As well, AI and ML can offer forecasts about how the future of manufacturing might look, they can’t explain how to make optimal decisions in that future – and what matters is making the right choices.
Quite simply it is because databases that include enough of the good and bad decisions to enable an ML algorithm to learn what to do in a given context just don’t exist. To create one would be a major challenge, not least because it would require a huge number of risky decision experiments to create such a database.
And in a year like 2020 where the things that have never happened before are wreaking havoc on global industries and making the right decisions in a period of disruption is more important than ever, being left blindsided by your AI and ML systems is a recipe for pain, if not financial disaster.
Manufacturing Analytics: Combining AI and Complex Systems Simulation
In much the same way that AI can complement humans, complex systems modeling and simulation can complement AI in manufacturing.
Here’s how we do just that at Cosmo Tech.
To start with, we leverage AI and ML for forecasting. AI algorithms are incredible at forecasting demand, for forecasting the aging of a machine, or for forecasting failures in a system. We can extract and leverage these probabilistic forecasts to provide inputs and insights, and we can feed these into our simulation engine. The simulation, in turn, helps us to organize our production to meet that demand.
Where the AI and ML manufacturing forecasts are data driven, the simulation engine is model driven. This model driven approach means that predicting things that haven’t happened before is possible.
More than that, though, as well as these predictive analytics, Cosmo Tech’s Simulation Digital Twins offer prescriptive analytics, too – think of these as extending the ‘what if’ capacity of prediction to the ‘how to’ level.
In a nutshell, it’s not just capable of telling what will happen next, it’s able to determine how to get to an optimal future. Hence, AI and ML provide the ‘what’ and simulation offers the ‘how’.
But importantly, the model driven simulation also feeds back into the AI.
The simulations, the results of the simulations, and the manufacturing analytics can feed into algorithms and help them to learn. Alongside the historical data that AI and ML algorithms rely on, the simulations provide the algorithms with access to future data. Thus, whether it is reinforcement learning, supervised learning, or unsupervised learning, data driven AI and ML approaches are complemented by the system driven simulation technologies, too.
AI and Simulation? Forget Either/Or.
Investing in AI and ML is something manufacturing businesses have been doing for a decade or more now. The quick wins were realized fast and encouraged further investment and expansion of optimization and forecasting programs, but they inevitably ran into the problems that all AI faces: with a fundamental reliance on historical data, an algorithm alone cannot predict what has never happened before or identify a causal relationship between data points.
Combine that data-driven AI or ML system with a model-driven simulation engine – something like Cosmo Tech’s Simulation Digital Twin – and you not only complement the investment in AI, you elevate it.
In the end, it’s not about an either/or choice.
You don’t just want a ‘black box’ ML algorithm that connects umbrellas with car accidents and recommends banning the former to save the latter. Instead, you want to choose both machine learning and simulation to better understand what will happen next and how to get to the future that holds the best outcomes for your business.