1. The origin of the hypothesis
By analyzing problematic processes, the experience of the workers, or the production line operation, the hypothetical probability is assumed on the possibility to improve or to change the process to increase outcome indicators (income, production output, sales, etc.).
So, the hypothesis of a process assumes that people themselves cannot physically take into account many factors and nuances at the same time since they are naturally inclined to cut the corners, assume, and work in their usual rhythm.
The hypothesis in machine learning is based on a greater amount of data when making decisions, which directly leads to a better-quality result in the end.
The advantages of ML with a well-developed hypothesis are in minimizing such human factors as injuries, stress, loss of concentration, etc.
2. Assessment of the hypothesis
A selection of the collected and initial data sets based on the hypothesis, as well as an assessment of the data suitability, is carried out together with the definition of the user and possible achievements for the subsequent integration of the model into the production process.
So, one should determine: for whom the results are created, who and how will use the results, what is the probability of achieving the desired indicators, based on the information collected.
3. ROI - economic effect and return on investment calculations
Specialists from the relevant departments (finance, efficiency, etc.) evaluate in cooperation of the economic effect.
This is what the hypothesis of the newly implemented solution was conceived and created for. Does it make sense to start new developments or projects? Will they be in demand and when it will be possible to count on profit and return on investment? At this stage, the metrics are identified (calculations of the number of potential customers, growth in the production output, expenses, and cost-effectiveness of consumables, etc.).
All parameters listed above form the goal that must be achieved.
4. The mathematical formulation of the problem
It is not enough to define and understand the necessary business results. The results must be converted into mathematical objects (graphs, tables, intersections) to define boundaries and dimensions beyond which the model cannot extend.
This stage is completed in collaboration with the customer so that the customer sets the limits or the thresholds (budget, upper and lower sales limits, product volumes, etc.).
5. Data collection and analysis
All information is collected from one place and then is analyzed with various statistical methods. Significant time is spent at this stage, but as a result, one obtains an understanding of the structure and the hidden relationships between various parts of the data for the accurate formation of the model features.
6. A prototype development
The essence of this paragraph in testing a hypothesis - to see if it works.
The model is built from the primary data of the results of the hypothesis testing. This is a simple and affordable way to find out if a problem can be solved or not.
The prototype clarifies the scope of the project and the possibilities for the solution implementation, and therefore finds its economic justification.
While a hypothesis is being created and the prototype is built, changes in the initial data may occur, such as a new product in the production line or new instrument in production equipment. In this case, the model should additionally be trained.
End-to-end processes and collaboration between different services with the Data Science team occur through DataOps and DevOps. This is very convenient, since they make corrections and additions at any stage, without interruption and loss of the result.
7. Creating a solution
Based on the results of the prototype, conclusions are drawn, and if they demonstrate good performance, then this is the first step to creating a solution.
The turnkey solution is being integrated into production. But to start, you will need to conduct training and retraining of employees, prepare equipment, etc.
8. Pilot and industrial operation
During the first few launches of the system, it should operate through the established test time in a team with a teacher-specialist.
This mode implies that the feedback between a human and the system will help to improve the accuracy of forecasts and make the necessary improvements to the system while normal operations are performed by a specialist.
If everything is fine then the test runs smoothly transfer into normal operation process. The next step is the transition to automatic maintenance and this is the final part.
The possibilities of machine learning of Artificial Intelligence are endless, but without the creation of models, they are not achievable.
And any model must be tested and verified before it is allowed to be introduced to operation.