Making sense of AI, machine learning & data science: A Different View
IBM estimated a few years ago that data growth rates have been so high that 90% of the data currently in existence has been created in the previous two years, with Tech companies and High-Tech Manufacturing particularly demonstrating the value that can be extracted from this flood of data. In doing so they have raised awareness in other industries that it is not only possible, but highly advantageous, to capture and process extremely large volumes of data. The catalysts that allow this value to be extracted from large volumes of data are Data Science & Artificial Intelligence. But in the same way that data volumes have exploded, so apparently have the number of tools available to deal with it!
At Hatch we find that our clients range widely in how mature they are in digital matters, and that they frequently have a lot of questions around exactly what AI can do for them, where they can obtain applications that use AI, and how they can integrate AI technologies into their organization to add value. Developing a clear understanding can be challenging as there are a huge number of tools, applications, platforms, languages and architectures available.
Another issue that further complicates understanding is that Data Science & AI technologies are frequently presented on a simple linear continuum, going from something like "Descriptive Analytics" up to "Autonomous Operations", or from "Threshold Monitoring" to "Digital Twins". The idea behind this representation is that as you become more digitally “mature”, so too does the complexity of the tools you use, and tools such as Descriptive Analytics are considered rudimentary counterparts to more “advanced” things like Autonomous Operations. There is some truth to this – the more you invest in Digital infrastructure, generally the more complex that infrastructure is going to be.
However, we find that a different way of thinking about the technologies under the combined umbrellas of AI & Data Science is much more useful: that they are a collection of inter-related tools that can each be used to solve a specific problem. They’re more akin to a toolbox than a roadmap.
The problems that these tools solve, fall into two main groups:
- Data reduction, insight generation & expertise capture
- Intelligent agents, simulation & optimisation
Digital Twins in a sense are the apotheosis of modelling: a representation of a physical asset that models all aspects of its behavior realistically and whose state is synchronized with the physical asset. In this article we’ll refer to Digital Twins to illustrate the inter-connectedness of the tools we’re discussing.
Data reduction, insight generation & expertise capture
These tools are most useful in the building of models, and they work by extracting patterns from data.
Manually building and validating models using Subject Matter Experts usually starts by applying a first-principles understanding of a system based on years of expertise. The result is usually very high quality, but this approach can be expensive as well. This limits the applicability of first-principle modelling techniques to the most essential areas of an operation.
In comparison, Machine Learning models can be rapidly created using only data. Machine Learning (ML), a subset of Artificial Intelligence, focuses on uncovering patterns in data. These types of models are valuable shortcuts. Given enough data we can quickly create a model of a system of interest. Tools such as AutoML or XGBoost can allow you to create high quality models almost automatically.
These automatically generated models can then be used to:
- Model the entire system (if the accuracy is sufficient).
- Approximate several small components that can be used in a larger simulation model.
- Help develop first-principles models, by providing insights into the behavior of the system.
In a very real sense, these ML algorithms learn about a system in the same way that a Subject Matter Expert does: by evaluating the available evidence and reaching conclusions about phenomenon X being associated with situation Y. A great deal of care is advisable in this area as Machine Learning algorithms have no significantly greater inherent capacity to differentiate correlation from causation, than humans do. Differentiating requires experimentation and the scientific method. This is part of the reason why we do not provide solutions without first validating them by a Subject Matter Expert.
Intelligent agents, simulation & optimisation
When you have captured the essence of a system in a model, there may still be questions that cannot be answered by the model itself. Consider chess: you have visibility of the entire board and knowledge of how all the pieces of the board act and interact, but armed with this information alone you still can’t answer the question of “what is the next best move?”.
The tools best applied to answer this kind of “what if” question are intelligent agents, simulation & optimisation. This group of tools focus on interacting with and controlling a system to determine the optimal outcome. As with the machine learning tools discussed above, the type of challenge you face determines the tools you need to use:
- If the system to be modelled is something relatively simple and straightforward we can use simple optimisation techniques (like Linear Programming or Stochastic Gradient Descent) to identify the optimum action.
- In more complex systems, simulation models such as Discrete Event Simulation can model a system repeatedly over and over to provide an indication of which actions have the highest probability of being optimal. This is a highly valuable application for a Digital Twin.
- If we have a complex system that changes over time (e.g. sensor calibrations drifting out of alignment) we can apply intelligent agents which not only make predictions of the optimal action but also learn and adapt to the response of the system. This response then can modify and guide the future behavior of the agent.
At this point you might be asking “So intelligent agents can learn? But I thought the capacity to learn falls into the other class of tools: Data reduction, insight generation & expertise capture”. And you would be right to be confused. This illustrates how effectively all of these tools all fit together and interact: intelligent agents can learn, and machine learning uses optimisation. As an example, consider work that Hatch has carried out in applying Reinforcement Learning (a type of self-teaching intelligent agent) to the challenge of efficient smelting:
Smelting operations are a challenging environment to optimise, as the smelting process follows a complex structure, and correctly managing performance is vital for production optimization and extending the longevity of process components. One of the fundamental issues that an intelligent agent must take into account is to balance between exploration and exploitation. In short, if agents are an “operator-in-a-box”, this operator must constantly choose whether to stay within the historical process parameters that previously resulted in optimum performance, or whether they should explore new process parameters that indicate that they may improve the overall performance of the system.
Multi-Objective Reinforcement Learning (MORL) is a subclass of AI which maximizes multiple rewards to solve optimization problems. This intelligent agent uses separate Neural Networks to learn from the data to (A) understand how to interact with the system, and (B) to predict how the system will behave. By causing these two networks to interact with each other, we have created an intelligent agent that can make optimal process control decisions.
The implementation of this work helps move smelting technology into Industrie 4.0, through data science, simulation, automation, IoT and cloud computing, for one of Hatch’s most non-linear, but data-rich process operation types.