Artificial Intelligence: Silver bullet for retail?

We recently asked food retailers what the most important technological trends are. 56% were convinced that artificial intelligence and machine learning will change the future of retail. Another 32% believe this to be true right now.

Artificial Intelligence, Machine Learning, Algorithms, Retail

AI Vs. Machine learning

But what exactly is machine learning? And how is it different to what we call artificial intelligence and the term deep learning frequently used today?

Artificial intelligence is the general term for a series of methods, tools and technologies that imitate human cognition, including recognition, learning, planning, reasoning and the ability to solve problems.

One of these methods is machine learning, aimed at making the computer capable of performing these cognitive skills. Machine learning is based on data and expert knowledge, out of which computers extract significant patterns and transfer these patterns to other data in order to generate predictions and recommendations. This method depends on computers identifying certain patterns or rules by themselves rather than being programmed to recognize them. Supermarkets use machine learning algorithms to plan optimally for fresh products and therefore achieve the highest availability while avoiding costly excess. A machine learning approach employs ‘neural networks’, whose basic structure reproduces that of neurons in the brain, albeit on a much smaller scale.

Deep learning is also about pattern recognition. It is based on large neural networks with many layers. It is often about things that a person can do well, but that are extremely difficult for a computer with rule-based programming: identifying objects in photos, understanding languages, steering a car independently. It is certainly not easy to train such networks, because in order to learn, they often require large quantities of data and substantial data centers with especially high computing power (GPUs).

In the end, though, all these different methods and concepts lead to one thing: software will increasingly be able to make quick smart decisions based on experiences made. Data is the sustenance – the more, the better. Depending on the problem, it will be data that is measured (historical) or obtained in a Monte Carlo simulation. There are also algorithms that create and keep optimizing complex strategies. All things considered, these are better at it than the best human. This is the case in games with clearly defined rules such as chess or Go. Here, the algorithms develop themselves so that they play against each other and, now and then, try something other than what they actually wanted to do.

Artificial Intelligence requires human capabilities

Programming artificial intelligence will still take a lot of expertise in the future despite all the advances. At Blue Yonder, we focus on our own algorithms (NeuroBayes, Cyclic Boosting). Our data scientists are continually developing these further to expand our technology portfolio wherever we see benefits for our customers while at the same time always considering current scientific research and methods.

Let’s go back to our example of retail product planning. Our machine learning technology is capable of predicting sales in the form of a probability distribution by using internal historical data (sales figures, stock information, prices) as well as external data such as weather, holidays or competitor prices.

Probability becomes reality

Scientists have proven that people cannot forecast the future with any degree of certainty in so-called complex systems (like commerce, economy, weather). The future is only determined once it arrives; in other words, when the future becomes the present.
Depending on the circumstances, sometimes it can be predicted accurately, sometimes less accurately. For every possible future, you can predict a probability. This kind of modern prognosis not only provides the median forecast value but also the complete probability distribution. Will it be sunny this weekend as predicted or will it unfortunately rain? Depending on which one, the BBQ might fall through. Do you know what a 60% chance of rain means in a weather report? It means: on 60% of the days with similar weather conditions as today, it rains. So ultimately this is about likelihoods, not certainties.

The quality of decisions depends on which costs and which benefits each possible future has. And that can be different for every article, as well as every location and time. Subjective customer dissatisfaction, sales price, purchase price and shelf life all play an important role in optimal product planning decisions in the face of predicted uncertainty. Both aspects play a role in creating value for the customer: the prediction as a probability distribution and optimizing the decision in this probability distribution.

The more valid historical data that exists, the better (customized) the predictions and decisions. Particularly in retail, where millions of decisions are required daily, the right balance between storage, pricing and sales is essential. The amount of available data allows for increasingly complex decision making. Therefore, it becomes more important for retailers to make automated decisions based on data, not only for product planning, but also for determining prices during normal operations and when selling seasonal goods, or when planning product ranges or for an individualized customer approach. This is where artificial intelligence is getting more important. The algorithms are constantly adapting to changes such as external influencing factors and therefore automatically making the best decisions. And this without any time lag – faster than a person can think, react and act. In light of the abundance of data and the speed of decisions, it is clear that the traditional gut feeling still used by many retailers to determine prices and plan for products just does not work anymore.

Humans are guiding the algorithms

If algorithms are better than people, will they replace us at some point? The discussion about intelligent software is intense at the moment, because it changes our working world and yes, it will take over some of the jobs we have today. This has happened before. When was the last time you counted your store revenue by hand? But our algorithms are here to support people, not replace them. They can take on routine tasks that are monotonous and partly also tedious, but that have become so complex, people can no longer manage them. The algorithms do not take away our responsibility to make the right decisions. In the future, we will also make decisions ourselves about the strategy for managing algorithms and any extreme exceptions.

Machine Learning: A Win-Win

In the discussion about digital transformation, it has often been said, but it is also true: the future success of companies is determined by the use of modern technologies. Machine learning is neither a miracle nor something terrible. It supports companies on a path into the future. To return to the subject of retail once again, algorithms see the future need for goods, can estimate the best price and ensure maximum margins for retailers. Thus the customer wins in the end as well: with the optimum supply of goods at the right time and at the best price. A win-win for everyone.

Further Reading:
For the study “Six reasons why retailers must make quicker decisions”, on behalf of Blue Yonder, the market research company Censuswide surveyed 750 decision makers from the USA, Great Britain, France and Germany. Censuswide followed the methods of the British Market Research Society (MSR), which are based on the principles of ESOMAR. You can find the complete white paper here.

Prof. Dr. Michael Feindt Prof. Dr. Michael Feindt

is the mind behind Blue Yonder. In the course of his many years of scientific research activity at CERN, he developed the NeuroBayes algorithm. Michael Feindt is a professor at the Karlsruhe Institute of Technology (KIT), Germany, and a lecturer at the Data Science Academy.