Following the "hype" about artificial intelligence (AI) that has been sparking such lively debate in recent years, we are starting to receive feedback of a serious nature on what machine learning can bring to business sectors offering on-site intervention services, notably in the area of problem-solving for resource provision and optimization. Before we look at an example of a typical application that I, personally, find very convincing, it will be useful to remind ourselves of the underlying condition that will decide whether or not an algorithm can yield results in terms of added value.
There is no such thing as Artificial Intelligence that ‘does it all’. Without data, AI can do nothing.
The promise of AI is the automation of a growing number of cognitive tasks – simple but repetitive or, conversely, complex – through deployment of algorithms loosely grouped under the title of machine learning (including deep learning), literally "machine learning". The specificity of these algorithms, and what sets them apart from traditional logical programmes, is their capacity to process cases for which they have not been specifically programmed. How can this happen? By "generalising" on the basis of source data that "trains" them during the learning phase.
There is nothing magic about this type of learning! It’s a simple question of statistics. A machine learning algorithm only learns from a large number of examples. It is within these learning data that it can identify invariables, correlations or recurrences that will enable, further down the line, «generalisation» to be applied, and so the correct processing – with a high probability of success – of hitherto unseen data. Whether it is a question of predicting a result, a quantity, a time and date, or to identify a specific behaviour, an object or a situation, the algorithm can only give satisfying results if it has been trained with pertinent data with regard to the problem considered, and data that is statistically representative of variations in the phenomenon one is seeking to model.
The important things to remember are:
- the learning process for a machine learning algorithm always requires availability of good quality data samples, in sufficient numbers. It is well-known that statistics derived from low volumes of data can result in misleading or erroneous generalisations…
- The error rate of this statistical type of learning statistical must be calculated at all costs, since it is this that will condition the reliability and effectiveness of the application envisaged, and even its relevance overall: if only 50% of predictions obtained for following the learning phase are correct, is the application really going to be of interest?
The example of spare parts stock localisation, optimized with AI*
Engie Home Service delivers for boilers, air conditioning systems and heat pumps to 1.5 million households and companies. The company manages:
- A geographic grid of 230 service points each with its own stock of spare parts;
- 300 technicians who perform 14,000 interventions every day, with 3,000 vehicles in which parts and equipment are stored as mobile stock.
Engie Home Service pledges repairs on demand for its customers. This requires expert management of and the provision of adequate competencies in the right place at the right time, but also stock control to ensure each vehicle has the right spare parts: to honour their promise to customers, the technician delivering the service must not only be able to diagnose any fault or breakdown, but also have access to the parts they might need in their vehicle. The problem is to anticipate which spare parts they must have to hand in each vehicle and at each geographic service point.
Moreover, because of the huge diversity of brands, models and types of equipment serviced, Engie Home Service has to manage a total of 1.4 million individual spare parts. At each service point there are on average 3,000 parts with individual reference numbers, and 150-200 in each service vehicle.
"When we take into account the sheer number of service points and vehicles involved, we arrive at a level of complexity that the human brain can no longer master" explains Stéphane Moillic, Director of Supply Chain d’Engie Home Service. And so, the idea came about of using Artificial Intelligence (machine learning) to anticipate requirements and quantify stocks needed, using as starting point historical data regarding movements of stocks and data for the installed base of boilers, with three objectives:
- reduce the value of stock holdings;
- increase the availability of spare parts;
- economise on transport costs.
After carrying out a survey on the quality of these data and a learning phase, a life-size test was carried out on stocking 9 branch offices over a period of 9 months, integrating all the relevant business constraints, and notably the variable of seasonal variation. The test yielded conclusive results on all three objectives, and so the model was validated and the experiment extended for a 6-month period to all the branch offices and service points, with the following results:
- 15% reduction in transport costs;
- 10% reduction in the overall value of stock held;
- 5% improvement in the availability of spare parts.
This is just for starters! At the conference where this case scenario was presented in March 2019, Engie Home Service set about launching phase 2 of their project, this being stock control and management for their 3,300 service vehicles, and announced a forthcoming phase 3 of the project to develop their network of ‘connected’ boilers: "using the data from our existing equipment base, we will be able to gradually increase our knowledge about which parts are likely to fail, and when, for one reason or another, explained Stéphane Moillic. In 2025, our technicians will have the spare parts they need in their vehicle before they even know they need them!"
An AI project can only succeed if it is accompanied by change management
In the first phase of their project, Engie Home Service had to given to the 230 staff managing the spare parts stocks in their branch offices. It was obviously impossible to just announce casually that "from now on, a machine is going to be doing your work in your place".
There is an anecdote that illustrates the type of fear that a project of this kind can arouse among operational staff quite well: when the algorithm forecasts the consumption of spare parts over 8-12 weeks, there are times when it will predict that there will be zero consumption of certain parts. It therefore puts the stock to zero for these items. A human being on the other hand, will reason that ‘zero’ presents a risk… We had to reassure staff by adding one spare part when this scenario arose, with the idea of checking later on to see whether this part had been used or not. In 95% of cases, it was the machine that got it right: in the event, there was no requirements for the part. This approach allowed staff to understand that the machine is above all a decisional aid, and that it could be tested along the way using their invaluable know-how, rather than dispossessing them outright of their competence.
When a human being has the time needed to study a case in depth, that person can often perform much better than the machine, but with the huge number of individual parts and more than 200 service points or branch offices, it is beyond the realms of the possible to study all case scenarios in depth. Once you move up a rung and confront so many individual cases on a wider scale, the machine will perform better than the human being. The human being, however, in all that is new or not routine.
In the case of Engie Home Service, the latest development is their completely new offering of boilers for a sales price of 1 euro. They have been able to make this pledge because their installed base is now under control, and managed by AI. Every mission has its own timeline: when the company wins a contract to maintain the boilers for an entire apartment block, of course they don’t immediately have statistics for the movement of stocks linked to these particular boilers: "It is human beings who have the competencies needed to manage these situations. Further down the line, after one or two months, the machine can take over. It is this switch between phases in every mission from now on that we absolutely have to keep track of and follow up with the right management decisions" says Stéphane Moillic by way of conclusion.
* Feedback [presented on 27 March 2019 in the framework of the SiTL