Managing inventory during and after crisis – Part 2 : Autoadaptation and collaboration with humans.

Ifmaking inventory decisions felt like casino-gambling during confinement, the after may be even harder… or not. Let’s share our experience – part 2 :

4. Considering Variations, Product Criticality and User collaboration :

Our Customers’ experience during this period confirmed that a proposed decision ( “how much should I order/move?” or “where should I allocate this item?”) should be blended with uncertainty and product criticality. This is particularly true for purely demand driven and buffer-oriented systems like Lean pull-system or DDMRP, or when working on short lead times, where variations take a greater role than forecasts at operational level. Indeed, in these cases, the inventory needed to cover the potential variation around the mean demand or mean lead time quite often exceeds the mean value itself.

Typically and to keep it simple, the idea is to adapt the answer to uncertainty based on product criticality (importance for customer to have it on time) and logistics cost for customer. If Nadii is able to help customers put a price on the impact of empty shelves depending on product type, the optimum coverage can even be precisely calculated.

Uncertainty management

Uncertainty and user collaboration :

This constant uncertainty monitoring is the base for real autonomy : having historical uncertainty, the system can :

• anticipate uncertainty (uncertainty can be partly anticipated, specially for seasonal products).
• react fast when spotting a sudden rise in uncertainty , and alert users before an accident happens – efficient collaboration with the users is  a key factor here . While auto-adapting to the changes, Nadii asked users for human confirmation of his decisions.  Since humans have access to knowledge that machines do not have, but they have limited time, experience shows that the system must keep the number of warnings under 1% to customer-side users to help focus attention on key issues.  There always be issues to be solved, just of different importance.

Nadii reports  to :

•  our supply chain data-scientists team for analysis, enabling the kind of response explained in strategy nr 1 (new feature engineering e.g.),
• users so that they control proposals. Nadii keeps a target to maximum 1% or alerts, usually way below, but a bit higher at the beginning of a crisis.

Pluses  :

• Aligns your teams, sales and supply chain, around a common money-driven business understanding.
• Performs optimised calculations that humans cannot timely perform (our customers could save double digits on overall logistics costs)
• Users feel secure while the system drives in autonomy and asks for help, without overwhelming them thanks to auto-adaptation.
• The security buffers are automatically adapted during the year, no need to steer manually

Minuses :

• Uncertainty is usually succesfully anticipated (seasonality…) , but for Covid-like events we must react rather than anticipate. At least reaction takes part of the pressure out of your teams that do not need to change parameters back and forth manually.
• Needs to define and maintain product criticity information for the system.

Some covid story

Once again the initial shock was not exactly anticipated, but the fact that Nadii adapts answer based on costs and criticality helped : decisions made before the crisis enabled customers to better protect important products from disruption.

While uncertainty rose, Nadii automatically suggested to gradually play safer to protect our customers’ critical high runners availability, while keeping logistics costs low for products accepting to delay customer orders. This still is reaction rather than anticipation, but at least takes part of the pressure out of your teams and optimises the money.

5. Machine-learning driven auto-adaptation :

Because one base of machine learning training is to learn from failure, Nadii measures accuracy of his decisions and archives years of results that he can compare to ground truth. And that is the base for auto-adaptation.

Just like a golfer would pick the right club and swing depending on the type of shot it needs to take, Nadii is able to pick the right algorithm mix out of a maintained base of algorithm and forecasts for each product, each family, for demand AND lead times.

These algorithms and associated features are fine-tuned for every particular customer, each being good at addressing a particular task (short/long history, low/high volume, features price and weather, …) for this particular customer, just like each club addresses a particular challenge. There are nonetheless 2 main differences with golfers :

1. Nadii will try all clubs, swings and whatever options on all positions before telling the right choice, and challenges the whole choices every week. Near-infinite training.
2. The second difference is that all the results are kept in Nadii’s memory, so that he can adapt the mix depending on the moment of the year for example. This is very useful for seasonal sales or around special events that repeat throughout the year.

It really makes a difference because of the incredible power of today’s computers, and if your find this strategy intelligent, you can call it AI. Our technical team prefers to use the term machine learning, since it’s more about intense training, but what we care about is that it is much more adaptable and robust compared to traditional approaches.

Without further being technical, here is aggregated answer of Nadii on the whole scope for one of our customers faced to great variations, in terms of lead-times and demand. This is comparison on a 30 k+ product portfolio, for the same customer, between Nadii and the traditional system customer used the year before.

Traditional statistical-forecasting based system

without lead time automation and auto-adaptation, year 1.

Nadii with demand and lead-times auto-adaptation.

same portfolio, +8% sales, -23% inventory, better availability, year 2.

One of the key factors explaining the difference above is that traditional systems ask users to set up manual security parameters, like lead times and security buffer, but in practice it just impossible to adapt them to changing conditions. This is just too much for humans to handle at each supplier and product level. Let auto-adaptation do it for you.

Pluses :

• Brings to users real autonomous driving, enabling your teams to focus on real added value.
• No need to manually tune parameters.
• Exceptional results in accuracy and robustness

Minuses :

• Requires the maintenance of a consequent database of algorithms and forecasts
• Requires to fine-tune the algorithm base for the customer products ( included in the set-up phase )