Listening to the expectations that people have about Artificial Intelligence (AI) and Machine Learning (ML), I am reminded of the old joke where a scientist is explaining to another a process and in the middle is a box with “Then A Miracle Occurs” written on it
People seem to believe that AI and ML will solve all their existing problems. I have been in many presentations where predictive algorithms and other machine learning tools are positioned to create a solution for the problem being discussed. Color me cynical, but I do not believe this. Yes, these tools will help with the solution, but people still need to be involved in implementing the solution (if there is one).
Let’s take a very simple maintenance example to illustrate my point. Both AI & ML can analyze data gathered about the operation of the equipment and learn to predict that there could be a failure in the future (commonly called predictive maintenance ) . With the result of this prediction, everything (in the presentation) is fine. In real life, the owner / operator now must do something with this prediction. Equipment must be adjusted or maintained, or the production schedule adjusted. If the course of maintenance is chosen, then staffing, operator skills, spare parts, purchasing time, must be considered. And if the lead time for the parts needed is sufficiently long enough then the prediction might be useless because the equipment is predicted to fail before the spares can arrive. In this case, the equipment could be run until failure. There is a plus to this, by running to failure you can certainly see if the predictive model is accurate.
Another “miracle” is the assumption that there is sufficient and relevant data to feed AI and ML so that they can generate the model that is being used for the prediction. I have been around enough companies that even getting various departments to agree on a simple calculation (with multiple data sources) is difficult. Identifying factors that can impact or influence the operation of and or maintenance of the equipment is another order of magnitude. I am also not convinced that companies currently are gathering the needed data to feed into the algorithms. AI & ML need a certain amount of historical data to learn from, and a lot of companies that I talk to do not have sufficient detailed data to feed such a process (e.g. some companies just collect the numbers of defects per production / process order, which is fine but does not lead to a very good predictive quality learning experience).
The Verge in 2016 pointed this out in an article “These are three of the biggest problems facing today’s AI”. The first problem they stated was “First you get the data, then you get the AI”.
And there is another problem, “another miracle”: the data needs to be clean. Ihab Ilyas, a professor in the David R. Cheriton School of Computer Science at the University of Waterloo said in his blog “When dealing with real-world data, dirty data is the norm rather than the exception”. And we know that Garbage In = Garbage Out. Having AI learn and predict from bad data is not to be desired. Thus, the data needs to be “cleansed”. Which is a complicated process with several issues needing to be resolved, including among others, missing values, outliers, data constraints, and cross field validation, Interestingly professor Ilyas suggests that “treating the data cleaning problem as a large-scale machine learning problem is principled and an obvious way to address these issues”
Don’t get me wrong, I believe there are opportunities for radical improvement utilizing the capabilities of AI and ML, but it is a long journey to utopia. I think that there is going to be a lot of work involved. And there will be problems along the way, after all, even with a lot of investment, autonomous cars are still having accidents (and sometime fatal ones).
So how to start your journey
Successful AI / ML projects must be driven by the business and should be designed to help to access and process data, and to facilitate decision making. To be successful there are some basic questions that need to be answered:
- What is the business problem or need that is to be addressed?
- What are the data sources available? Do you need to find additional sources?
- Are the data sources sufficient & do you understand the meaning of the data in them? Has the data been “cleansed”?
- Is the data is going to be used in training the model representative enough? Do you have enough data to “test” the model?
- Are there any biases that are not factored out? Can you even identify the biases?
- What algorithm are you going to consider and train?
Your project can be successful. Identify a concrete problem and go forward. After all, we are not trying to change the entire universe, just our little part of it. As an example: Trenitalia uses SAP HANA Platform to analyze data coming from thousands of sensors aboard their trains in real-time, helping them improve maintenance processes and ultimately offer better customer service.
Of course, it is even better when you do not have to create anything. Commercial software vendors are starting to embed AI / ML into various business processes. Booking travel and expense reporting is a common business activity for many companies. Customers who utilize the SAP Concur solution are already using AI /ML. Even if they did not know it.
Another area where SAP is utilizing AI / ML in a standard business process is in the finance and accounting operations. New incoming payment and open invoice information is passed to the SAP Cash application, a cloud based matching engine, and the proposed matches are either automatically cleared or suggested for review by accounts receivable area for processing.
As you can see, AI / ML is already happening. You can be using it today without your knowledge.
What has your company’s journey been with AI or ML? Can you relate? Please share your thoughts and ideas.