Amazon’s Drone delivery approval and Ants- What is the connection ?
So you have probably already read the news, in many different mediums regarding FAA approval of drone deliveries for Amazon. Now that Amazon has checked the box on an easy action items, the next step will be to test the real execution. Drone deliveries, from a planning and execution perspective, are much much more challenging than routing ground transportation, and even conventional air transportation.
Why ? Trucks have defined paths in terms of roads, each road named and marked. Even conventional Air traffic has air lanes defined. But that will not be case for drones. Now initially there may not be that many drones in the air. But once these deliveries become usual and frequent, hundreds (or manybe thousands) of such drones will be hovering, in “swarms” over a city. What type of Algorithms can help ensure they do not run into each other, into other high rise features and yet find the quickest or shortest path to the customer ? I already dropped a hint when I used the keyword “swarm” in one of the preceeding sentences.
Smart ants and their swarm intelligence
Ok…so it is generic information that ants, generally, move in swarms. Now here is an experiment that you SHOULD try, if you can find an ant colony in your backyard (I already did,). Place a sugar cube 2 feet away from the ant colony, within the range of scout ants around the colony entrance(s). Come back after 30 minutes. You will see that ants are converging on that sugar cube from many different directions, working diligently on breaking it into tiny particles and carrying it back to their colony.
Come back after ~ 1 hours. You will see that ants are now following a single line to haul their loot back. And here is the scary part- You will realize that as compared to several other sparse lines that you saw earlier, this dense line is the shortest path back to the colony.
Ant Colony Optimization: Consistent experiments (done more sophisticatedly than mine or yours), show that ants always converged to the shortest path between their nest and the food source. This outcome has been consistently observed in asymmetric bridge experiment with real ants. In controlled lab experiments, this convergence can happen in less than 10 minutes.
So yeah, these ants are performing an optimization of their own. Given the general understanding of swarm intelligence in ants, scientists have developed a heuristics popularly known as Ant Colony optimization, that helps solve many problems, when several constraints need to be met in a specific problem space and an absolute best solution is difficult to find due to a vast number of possible solutions.
And in the case of these drones, these swarm intelligence optimization methods will be the ones that can come handy- powered by massive computing infrastructure that can perform dynamic route optimization for these drones in near real time.
But I was going to talk about Heuristics- wasn’t I ?
Thanks for your patience so far. The real discussion starts now. Swarm intelligence algorithms (like the Ant Colony Optimization algorithms) that can help route drones , fall in the category of what is known as Heuristics algorithms. Let us start with a standard definition of a Heuristics
“Heuristics are a problem-solving method that uses shortcuts to produce good-enough solutions given a limited time frame or deadline. Heuristics are a flexibility technique for quick decisions, particularly when working with complex data. Decisions made using an heuristic approach may not necessarily be optimal”
May not be optimal ? Yes- because the vast number of possibilities mean that it is next to impossible to calculate the optimal method. But, a key aspect of the above definition, that is more important for me is:
“Heuristics are a flexibility technique for quick decisions, particularly when working with complex data.“
And this- not the complexity of problem- is the reason I think that Heuristics will define the future of Smart Supply Chains. Ants find simple solutions to complex problems-and that is the part I love about using heuristics.
My Quest for that “Holy Grail” Algorithm
I have been trying to “self teach” myself some of the advanced Machine Learning (ML) for a while and then I thought that I needed more structured around it. In my quest to find applications of ML in Supply Chain domain, I was either finding either basic applications (classification, clustering etc.) or very advanced (like Deep learning), but nothing in between. Traditional OR optimization algorithms were the only ones that I could think of in between these two categories- no “pure” ML algorithms.
So I though I should probably enroll in a program that covers the gamut of algorithms more extensively and in a more structured way. I enrolled in a program and at this point, have already covered most of the algorithms other than NLP, Deep Learning and Reinforcement learning. I was still not able to find applications beyond regression. clustering, classification- many of which have existed for a while.
And then I realized that the way I was thinking was flawed !
Developing the new kind of ” Ensemble”
And then I realized the “missing link”. And it was funny that it took me a while. Because in one of my jobs, I developed a Heuristics- a logic, an algorithm, coded around unique operating aspects, to optimize schedule of a day to day operations aspect (can’t share more details due to confidentiality reasons).
In Machine Learning parlance- Ensemble algorithms are algorithms that leverage more than one Machine learning algorithms. But what we will need in the world of Supply Chains, to fill the gap between applications of basic Machine learning algorithms and advanced Deep learning methods, is a new type of “Ensemble” algorithms that combine heuristics with classic ML algorithms.
Supply Chain processes, where most of them stand right now, even for best in class companies, are very fluid from one company to another. One size does not fit all in the world of Supply Chain. Forget about different companies, even two DCs within the same company have different ways of doing the same thing.
And that is why when I tried applying ML algorithms, as they exist, I either could not leverage it successfully or saw dismal results.
I initially thought that it was my Frankeistein codes (cobbled together from different pieces from internet) so I hired developers twice in 2019 and early 2020, only to find that though my code definitely needed improvements in other aspects, the results were ok.
Extreme customization of Supply Chain processes, fluid nature, deviances and several other aspects mean that no one standsrd algorithm can effectively capture a complex proces in its entierity. We can force fit something but it will not do us any good if we really want to use ML to get innovative solutions, not just to “show” that ML is being used.
They key is to build Heuristics- algorithms defines around the process nuances and embedd ML algorithms therein, thereby creating a solution that will meet the exact requirements.
Another great aspect of this approach, in addition to the fact that you will have an algorithm to define a solution that currently does not exist off the shelf or in a synadardized algorithm, is that as the rules around your process evolves, you can make tweaks in your heuristics, unlike the algorithms built using standard libraries, where customization is not impossible, but no as intuitive. In fact, you can build a decent UX on your solution where parameters and rules can be played with by your process managers.
The “Smart” Supply Chains can be powered only by leveraging Heuristics as foundational algorithms !
Views expressed are my own.