The history of Artificial Intelligence goes all the way back to the classical antiquity where the idea of intelligent robots (such as Talos) and artificial beings, such as Galatea and Pandora, were a part of Greek Mythology (McCorduck, P., 2004). Talos was a giant automatic machine circling the shores and protecting Europe from pirates and invaders.
But even if the idea and concept of artificial intelligence and robots is old, we are still struggling to live up to and achieve the ideas of our ancestors. The major achievements within AI have been enabled by computers and the exponential growth of computer power and storage capacity we have seen over the last decades.
Using AI today
We can define artificial intelligence like this: “Artificial Intelligence (Al) is the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages,” (Techopedia, 2020).
Our main purpose today is to look into different ways of using AI to optimize performance of processes within production, in addition to traditional APC (Advanced Process Control) techniques. APC covers a broad range of techniques and control methodologies, such as fuzzy logic and statistical control.
The most relevant branches of AI that may be used for optimizing processes can be classified in two different approaches:
- Prior knowledge (deterministic rules), including knowledge-based rules and logical systems.
- Data driven or data analytics approach, where we use machine learning and deep learning techniques.
The challenges with a prior knowledge approach
The domain experts are extremely important when establishing prior knowledge approaches. The domain experts are used to set up the rules for the knowledge-based systems. However, the challenges are always the same when people are involved:
- Experience of domain experts
- First time/rare occurrence (learning)
- Cognitive bias (pre-conceived notions/lock-in)
- Knowledge management (Access to knowledge)
- Distinction between beliefs and truths
- Cross-discipline expertise and the whole organizations experience
The challenges with a data-driven approach
Data analytics and machine learning is a good approach to use on large data sets to learn from past observations. This approach is attractive because we can have a “black box” approach where we, in an ideal world, feed it with data and pull out the results. In the real world, there are a few challenges for this approach as well, including:
- Quality of data (raw data are a combination of noise + data)
- Need for humans to validate data
- The model needs to be trained for the events we design to avoid
- Testing the model involves humans
- The results are difficult to explain to the users
Hybrid ML/Analytical models
Some areas have seen great results of machine learning, despite the challenges listed above. Hybrid models, where machine learning is combined with prior knowledge analytical models, is currently being investigated. One way of doing this is to add synthetic data generated by prior knowledge models to mitigate some of the drawbacks of machine learning on its own. Hybrid Machine Learning/Analytical Models for Performance Prediction is still early days, but an area to follow closely in the future.
Human Intelligence vs. Artificial Intelligence
If the implementation of artificial intelligence is still heavily dependent on human intelligence, this should be used from the start. Is there a way we can digitalize process knowledge? A good approach is to reduce the reliance on historic data and enhance the role of explicit rules (knowledge graphs) generated in a deterministic scientific approach. Letting the users participate in the validation and enhancement of the models with implicit experience built over years of operation.
This is the goal of a branch of AI called Qualitative Physics and more specifically a part of Multi Level Flow Modelling (MFM).
A few important observations:
- Today, all AI applications require some human interaction during set-up or validation.
- Machine learning alone cannot predict incidents that have not yet occurred.
- For a successful AI deployment, explainability of the model is vital.
- McCorduck, Pamela (2004), ‘Machines Who Think (2nd ed.)’, Natick, MA: A. K. Peters, Ltd., ISBN 978-1-56881-205-2, (Accessed 20 Feb. 2020).
- 'Artificial Intelligence (AI)', Techopedia. Available at: https://www.techopedia.com/definition/190/artificial-intelligence-ai, (Accessed 20 Feb. 2020).