Can machine learning save the planet?

Dr Gah-Yi Ban explains that inventory theory combined with Big Data could help address the problem of global waste


Each year, one-third of food produced globally for human consumption is lost or wasted globally, according to United Nations estimates. That amounts to about 1.3 billion tons. While most of us are guilty of throwing away left-overs at home, the bigger culprit is actually at the supply side: the US Department of Agriculture calculated that in 2010, some 30-40 percent of the food supply was wasted before it even got to the plates of American consumers.

The waste problem is not limited to food. The True Cost, a 2015 documentary film about the human and environmental costs of fast fashion, estimates 11 million tons of textile waste are generated every year in the U.S alone. A United Nations Environment Program report estimates up to 50 million tons of electronic waste will be dumped in 2017.

Most of us are aware of the enormous environmental cost of waste production, and do what we can by recycling and careful purchasing. Urging individual consumers to play their part is one important piece to the global waste challenge. But an important – although less well-known – flip side to waste reduction is smart inventory management: in other words, not producing more than is wanted in the first place.

Setting the targets

The mathematical theory of inventory management is well-developed. It can be traced back to giant names in academia who worked on the problem in the 1950s, including Kenneth Arrow, who passed away recently and remains the youngest-ever winner of the Nobel prize for economics. A second key figure was Herbert Scarf, one of the founding fathers of operations research. Early works, led largely by Scarf, focused largely on the structure of optimal inventory management. Through ingenious mathematical insight, Scarf was able to show that the optimal way to manage inventories is through just two critical numbers, let’s call them s and S (little s and big S). Put simply, if the stock of a good falls to s or below, then the business – say, a newsagent – should order sufficient papers to bring the stock up to S. Put another way, the size of s determines at what stage an order should be placed, and S determines the size of that order. Subsequent works by Scarf and other researchers have found that the structure of the optimal inventory policy remains of this (s,S) type even in much more general settings.

Big data, big opportunities

Now, more than half a century since the pioneering work of Arrow and Scarf, we are moving into the era of “Big Data”. In a series of recent papers, my co-authors and I have shown how companies can use raw data to compute the optimal inventory policies in a variety of settings. In particular, we demonstrate that data-driven inventory algorithms can reduce the total cost of the operation by up to 24% compared to best practice benchmarks. The cost reduction is a result of better matching of supply and demand – in other words, by not producing unnecessary goods when the demand is low and producing enough when the demand is high.

Once the link from raw data to the optimal decision – in this case, the optimal inventory policy – is found, such decisions can be automated. Although machine learning and automation have been receiving negative coverage with doomsday warnings, smart inventory management is one domain where automation should be welcomed. This is because optimal inventory management is inherently a complex mathematical problem, one that the unaided human brain cannot solve. Just as we let algorithms to search and rank from the entire internet; complex ordering decisions should be designated to smart algorithms.

In short, it is now possible for companies to convert raw data to optimal inventory decisions. With benefits to both the bottom line and the environment, time is ripe for industry-academic collaboration in smart inventory management.

1. Scarf, Herbert, The optimality of (s, S) policies in the dynamic inventory problem (1959). Mathematical Methods in the Social Science, KJ Arrow, S. Karlin, P. Suppes, eds.. Stanford University Press, Stanford.
2. Ban , Gah‐Yi and Rudin, Cynthia, The Big Data Newsvendor: Practical Insights from Machine Learning (February 6, 2014). LBS working paper. Available at SSRN:
3. Ban , Gah‐Yi, The Data-Driven (s, S) Policy: Why You Can Have Confidence in Censored Demand Data (August 31, 2015). LBS working paper. Available at SSRN:
4. Ban , Gah‐Yi and Gallien, Jérémie and Mersereau, Adam, Dynamic Procurement of New Products with Covariate Information: The Residual Tree Method (February 24, 2017). LBS working paper. Available at SSRN:


Comments (2)

cagbakwuru978 5 years, 2 months and 18 days ago

Machine learning is a new technology derived from Artificial Intelligence performance network which dont agree to a large extent with the natural realities both scientifically and traditionally, there products of machine learning mechanisms cannot be tenable in applications of the former in the latter because of non conformity principles of law of nature. Machine learning in the foregoing case study applications will not produce any reliable solution to the problems of global waste.

cagbakwuru978 5 years, 6 months and 29 days ago

If machine learning means the use of modern machinery to proceed with scientific research methodologies that provide clues and solutions to Global problems, why in the contrary.