WarehouZe – From Silicon to Carbon for Gen Z

In a not too distant future, as you sit on board your next airplane on the parking bay, comfortably enjoying your fresh orange juice and hot towel as you sink deeper into your game changer seat. It's on-time departure could be delivered via Deep Maintenance Dispatching (DMD), the name is inspired by the paper referenced below.

How can we make aircraft maintenance facilities more efficient? One approach could be to make them more intelligent. That's what researchers at Hitachi America Ltd are trying to do as detailed in their new research paper where they improve dispatching systems in (simulated) warehouses via the use of Artificial Intelligence. They call their methodology "Deep Manufacturing Dispatching (DMD)", which if you ask me is a very cool name and something which I believe is very applicable to aircraft maintenance in equal measure as it is to manufacturing outfits.

How does heuristic optimisation work? A heuristic is a technique designed for solving a problem more quickly when traditional methods are too slow, or for finding an approximate solution when classical methods fail to find any exact solution. This is achieved by a trade-off, sacrificing a bit of optimality, completeness, accuracy, or precision for speed. In a way, it can be considered a shortcut. We normally use heuristic optimisation methods to find "good enough" solutions quickly and these are used in solutions like INFORM's Groundstar suite, IBS Airline Solutions and Quintiq's suite of solutions, to name a few that I am quite familiar with.

So traditionally warehouse picking and delivery queues are handled through this heuristic optimisation software. The downside, these are very intricate and involved software with a lot of tweaks, knobs and tricks up their sleeves and hence there are a handful of companies in the world that do this really well and sometimes even with them the end result doesn't work out as desired.

How does this DMD thing do its do? The researchers turn the state, a snapshot of the status of things that you are concerned about, on the shop floor into a 2D matrix, incorporate various parts of state from the environment, then design reward systems which favour the on-time delivery of items, be that in the hangar or out on the line or in the shop. This concept generally in the airline industry, applies to a long list of problems like people (Resource Allocation), aircraft parts delivery (Stores), Cargo, Ground Support Equipment or Airplane tail assignment just to name a few, take the state and environment and convert that into a set of numbers and then design a carrot and stick system that rewards the correct behaviour and punishes the bad ones, it’s almost like Pavlov doing his dog thing!

Does any of this stuff really work? Surprisingly it actually does and rather well. In a simulated environment: The researchers compared DND results with seven other dispatching algorithms covering the entire range, from carefully crafted rule-based systems, to ones that use ML and Reinforcement Learning. The team performed these comparisons in a range of circumstances, observing how well DMD could satisfy different constraints – specifically in the case of part supply, lateness and tardiness. "Overall, for 19 scenarios, DMD got best results for 18 on discounted total reward and in 16 on mean lateness and tardiness." In other words in these monitored assessments, DMD whupped other systems backside …… repeatedly.

It might be uber-cool but why does this really matter? As our increasingly digitally-driven economy becomes even more digitised, we can expect parts of the physical goods chain to move faster, as some goods are a function of people's likes and dislikes which are themselves determined by social media and “viral” digital things. Papers like this highlight the point that in the not so distant future the choice of offerings would feed on itself and become even larger with more retailers offering an even larger variety of products. A sample size of a hyper-personalised handful, each Stock Keeping Unit (SKU) sold at previously unheard of low volumes, all while increasing the importance of smart self-learning digital systems for efficiently coordinating warehouses. In a world where every individual wants to be unique, how do you make money as a retailer without economies of scale, the answer is by doing things like this and many many others to lower your costs.

The renting economy is taking over the world. The only thing standing in the way of Prime Air from Amazon and our industry, in general, are the regulations and as the youth of today are the policymakers of tomorrow, I strongly believe a time is coming when airlines would operate as subscription services much akin to Netflix where you buy flying KMs that come bundled with other services like Amazon Prime does and not buy tickets for city pairs. We saw this kind of swift change in the US when the 1978 deregulation happened and the US airline industry in the 1980s was dramatically different from that of the 70s. Imagine economic aviation deregulation of the entire world where bilaterals and slots are not an aeropolitical discussion but an economic one. The entire industry would change and I for one believe quite strongly that such a time is nigh.

As we move into an era of mobility as a service that is hyper price-sensitive, mission and operating equipment matching would become even more important and a laser focus on costs is here to stay for the foreseeable future. Since airplanes have a lot of SKUs especially if you operate a multi-type fleet, anything that would help us improve on the costs front would become crucial. I see reinforcement learning as an autonomous and continuously improving policy learning and operations management vehicle, which would run many parts of the airline of the future. We are not there yet but I believe a time is coming fast where operations planning, network control, maintenance scheduling, disruption management, stand planning, warehouse management and many other areas of the same ilk would have significant pieces of reinforcement and transfer learning embedded within.

What about the environment? So Reinforcement learning is important but where does the Transfer Learning part come in, let me explain. Environmental impact of aviation is going to become even more of a concern as the skies above Europe keep getting crowded not to mention initiatives like CORSIA would start adding new dimensions to the costs of running an airline. Carbon footprint of organisations is going to really start hurting the purses sooner than we think and given the climate impact being green, sooner or later, will be the only way going forward, as it should be. A tonne of Jet-A1 fuel roughly produces three tonnes of CO2 which means given the spot rate, you will not only be paying for the tonne of fuel but for three tonnes of emissions too. AI too has a carbon footprint as it uses large computers in large data centres that run using electricity. The world is abuzz with the things that AI and ML can do but its carbon footprint is something people do talk about but did not really know how much. New estimates coming out of the University of Massachusetts Amherst in the US suggest that the carbon footprint of training a single AI is as much as 284 tonnes of carbon dioxide equivalent – five times the lifetime emissions of an average car and this is why I believe that transfer learning would become more popular. So what is transfer learning? Transfer learning is a process of making use of the knowledge gained while solving one problem and applying it to a different but related problem. For example, the knowledge gained while learning to recognize cars can be somewhat be used to recognize trucks, which means shorter training time for the new task and leading to a lower energy requirement and a smaller carbon footprint.

I have been a big fan of reinforcement learning for years and have advocated its use internally with a fair bit of fervour on how we could use it in an airline operations and disruption recovery setting but to be honest the tech wasn't ready for primetime use but things might be changing and I can't wait to see what happens next.