Machine Learning and Automating Decisions - Algorithms Define Action

IN Machine Learning — 17 March, 2017

Machine Learning and Artificial Intelligence (read previous blog article) are poised to disrupt industries one by one. In fact, early adopters in retail are already seeing substantial benefits from using machine learning and artificial intelligence in their daily operations in areas such as replenishment and price optimization. Peter Sondergaard, Senior Vice President at Gartner and Global Head of Research at Gartner Inc. points out that: “Algorithms are where the real value lies. Algorithms define action.” He also adds that “People will trust software that thinks and acts for them”.

Machine Learning and Automating Decisions

Humans making decisions

What does this mean for a company’s daily operational decisions? Depending on the size of the business, the range of products or services offered, how many customers need to be served each day, a large number of decisions need to be taken on a daily basis. Half a billion decisions per day may seem an overestimation but is quite realistic for something like a supermarket or drugstore chain with tens of thousands of products across hundreds of stores that need to be ordered on a daily basis – including projections into the near and intermediate future. Traditionally, all business decisions are made by people using their expertise to make the best decision in a given situation and then take appropriate action. However, in scenarios like that of the grocery chain, the sheer number of decisions alone make it prohibitively difficult to allocate the appropriate time needed for each decision.

And what if we had the time to deliberate each decision until we’re convinced we’ve made the “right decision”? Behavioural research shows that humans are not really good at making decisions. From an evolutionary perspective, we make decisions as a way to identify threats: We analyze our sense of normality to identify threats. We look for causal relationship as a way to make sense of the world with the information we have at hand. This has guaranteed our survival as a species,

However, in today’s more complicated world, decisions are beyond mere survival. Decisions come from a myriad of contributing factors and our natural way of processing information leads to cognitive and social biases that can influence our decisions. An extensive list of such biases can be found on Wikipedia, Business Insider has also compiled a handy chart of the 20 most prevalent biases.

What we humans suffer most from is overconfidence. Experts in a given field tend to lean on their expertise and overestimate their own performance or show unwarranted certainty in their decisions. Indeed, the Nobel laureate D. Kahneman said in an interview with the Guardian that if he could change one thing about humans, it would be overconfidence. Further, people don’t have a good intuitive understanding of probabilities and stochastic effects: As our day-to-day experience is mainly guided by deterministic behavior, it’s very difficult for us to think in probabilities and associated uncertainties. An academic study by G. Bolton and E. Katok (Preprint) showed that individuals are prone to “chase” after short-term fluctuations. The study asked participants to make a “daily” decision to determine how many perishable articles should be ordered to satisfy demand and still turn a profit. The condition was that all goods expire at the end of the business day, the so-called “ultra-fresh” in the retail domain. In each turn of the computer-based study, the participants had to specify an order quantity, after which the simulated demand and realized profit were revealed. The setup was then changed such that the participants were forced to place orders for the next 10 periods (instead of one), which effectively forced the participants to take a more representative sampling into account and impeded short-term demand chasing. The authors of the study found that restricting the participant’s ability to interact with the system led to significant improvement in the overall performance. In a way, moving to standing orders is a crude first step in automating decisions, as the human expert is committed to let the system operate by itself, although some provisions can be made to interact in the event special circumstances require manual interventions.

Predictive Applications & Machine Learning

A better way would be to design, implement and test a system that can do most of the decisions we need to make efficiently, without bias and at scale, and then optimize each decision instead of just relying on periodically adopted standing orders.

Predictive Applications and Machine Learning

This brings us back to the “algorithmic economy” Mr Sondergaard discussed earlier. “Predictive applications” allow for automation of operational decisions using machine learning and artificial intelligence as shown in the diagram: All relevant data are transmitted to the predictive application, which ingests the data and checks for inconsistencies. Sophisticated machine learning algorithms then predict the future behavior and derive the optimal decision while taking all relevant constraints into account. External information, such as weather forecasts, competitor information, etc., can be included as needed. The optimal decision is then fed back into the operational system, which sends it to the execution layer. A business expert has access to all relevant data and decisions derived from the data and is able to monitor the performance of the predictive application and take action if needed. Can all decisions be automated? No, but around 99% of all operational decisions can be. However, we know it’s important to be able to take control if the need arises, such as a crisis or extreme circumstances the machine has no knowledge of. Automating 99% of all operational decisions allows the experts to focus on the remaining events where human expertise and judgement is needed. Coming back to the 500 million decisions that need to be taken at a specific retailer on a daily basis this means that 5 million decisions still need to be checked by a person each day – which is still a very large number. However, quoting D. Kahneman again: “The prejudice against algorithms is magnified when the decisions are consequential.”

An example from the real world

Prescriptive Analytics and out-of-stock rates

How does it work in practice? The following example illustrates the benefit of a replenishment system powered by sophisticated machine learning algorithms can bring to a business: A large German supermarket chain was plagued by high out-of-stock rates of around 7 %. Initially, Blue Yonder Replenishment Optimization solution was used to deliver the best decision as a recommendation to the local experts , who were then either able to follow the daily recommendation or change the suggested order quantity according to what they thought was best (“prescriptive analytics”). The out-of-stock rate dropped to around 5 % without increasing inventory levels, which was already a significant improvement over the previous situation. In the final step, the system was switched to automated orders for almost all operational decisions, which implied that the retailer’s experts were now free to get more deeply involved with special circumstances that the machine learning system has no direct knowledge for. As a result, the observed stock-out rate dropped to around 1 percent, again without increasing inventory levels.
Moving from prescriptive analytics to automated decisions, putting trust in the machine, improved the business substantially, or to quote Gartner’s Mr. Sondergaard again: “People will trust software that thinks and acts for them”.

Dr. Ulrich Kerzel Dr. Ulrich Kerzel

earned his PhD under Professor Dr Feindt at the US Fermi National Laboratory and at that time made a considerable contribution to core technology of NeuroBayes. He continued this work as a Research Fellow at CERN before he came to Blue Yonder as a Principal Data Scientist.