What’s the story?

Arriving at work this Monday morning, people at the building I work in encountered four security people eager to assist them in using the new elevator system that was put in place over the last week. The excitement was palpable and the anticipation could be smelled in the air.

As someone mildly obsessed with elevators and their (dys)functioning I was particularly looking forward to this new experience. I approached the first man, notifying him that I would like to be elevated to the 29th floor. Upon receiving this information, he regretfully informed me that he would not be able to be of service to me and requested I continue to the next person along.

Background

To understand what just happened, one has to have some knowledge about our set-up. Due to the tapering shape of the building, as well as presumably for efficiency reasons, there are two sets of elevators servicing the building. The first set of six elevators stops only in the lower half (Ground – 17) and the second set of six only in the upper half (17-34) in addition to the Ground and Mezzanine floors, the latter containing the canteen.

This could work rather nicely, separating out passengers, were it not for the frequent requirement of people from the higher floors to visit the lower floors and vice versa, requiring a commute that involves catching a connecting elevator on the 17th. Furthermore, to ensure it doesn’t become too simple, half of each of the two sets of six elevators (are you with me?) has the ability to go down to the lower ground floor and basement levels, containing the auditorium, whereas the other half doesn’t. Seeing as you can only indicate that you’re going down but not where to, you have a fifty percent chance of avoiding another connection when making your way to the basement.

It’s ok. We can handle this. Were it not for the fact that, for some inexplicable reason, frequently only one of each set of elevators seems to be doing all the hard lifting, literally. On numerous occasions people have been found awaiting their vertical carriage after lunch only for a completely full elevator to teasingly open its doors and reveal its sardinesque contents. Again, this is manageable… until the doors close again and the indicator above the door reveals that this will in fact be the next available elevator (!!). Somehow, this completely full elevator, en route to dropping off its inhabitants on a scattering of floors between 17 and 34, and -presumably- picking up people on the way down, is deemed to be the best possible elevator to attend to these worker bees, keen to get back to their daily grind. For some reason, the five remaining elevators in the set are on their lunch break, on strike, or auditioning for a role in The Shining or –the equally recommendable- De Lift.

A new hope

So, finally, after many years, a wonderful, magical, new solution has been put in place. Shiny new displays in the elevator lobbies allow passengers to select their desired destination and the screen will inform them which elevator will be transporting them there. Surely this would make everything better. Now, rather than only knowing a traveler’s intentions once inside an elevator, the system has advance knowledge. It can start using this valuable information to intelligently decide how to distribute its passengers as quickly and efficiently as possible to their respective floors.

Unfortunately, as we’ve already witnessed the first indications of on that fateful first morning, elevatorial bliss appears once again to be eluding us. Surely, (surely!), it would have made sense to design and implement a system smart enough so that the user, including visitors unfamiliar with the taper-induced intricacies of the construction, need not have to work out which set of screen devices to use! Surely, that would be design principle number one for a system designed to inform people which elevator to use.. the ability to inform people of which elevator to use!

Later that day, I decided to test if indeed it would not be possible, and proceeded to enter 29 in one of the screens next to the lower half elevators. The worst that could happen would surely be an error message? Nope. After receiving the first digit, the system was smart enough to work out that my only possible destination would be floor 2 and smugly informed me that elevator B would shortly be arriving to take me 7% of the way to where I wanted to go…

Requesting an elevator on the 24th to go to the 8th? “Sorry, never heard of 8th mate”.

But at least passengers will end up smartly clustered based on their desired destination, right? Right, smartly all clustered into the same elevator with the doors tentatively closing, only to reopen at the last minute as the next person has finished entering in their destination… Sigh.

On the history of elevator control design strategies

This isn’t my first run-in with evil elevators and I’m sure many of you can identify with the above frustrations. It has always amazed me how it can be possible that something that’s both so obvious a logistical challenge and so frequently a nuisance can have gone unresolved for so long.

Cue a dive into the fascinating literature of elevator control design systems.

The first pioneers in the field of elevator traffic published their work in the early 1970s, including G. D. Closs’ dissertation at the University of Manchester. It however wasn’t until the publication of Strakosch’s The Vertical Transportation Handbook in 1988 (buy here) that the mainstream public took proper note of the advances happening behind closed doors, again, literally.

I hear you wondering whether this is the same Strakosch as the inventor of the zone approaching strategy and, yes, it is of course. For those of you who haven’t recently brushed up on your classical elevator control design strategies, the zone approaching strategy is where the building is divided into a number of vertical zones, with elevators assigned to specific zones. If there’s nothing going on in their zone, they can just hang around (lit.) for a bit.

Contrast this with the other classical strategy of collective control, as devised by Siikonen in 1993 in her seminal work, Elevator Traffic Simulation. Here, elevators rush to the nearest call request in their current direction, competing with each other as much as cooperating. If Strakosch is the man behind zonal marking, Siikonen is the lady behind man marking.

These heuristic-driven elevator control design systems however are both difficult and costly to design. And surely, elevators lend themselves perfectly for adaptive learning methods? So what advanced have been made in the field of artificial intelligence and machine learning in particular? It turns out, quite a lot.

Artificial intelligence, machine learning, the internet of things and elevators

Elevator systems should be the ideal hunting ground for machine learning enthusiasts. The non-stationary nature of the problem to be solved, with different situations presenting themselves at varying times of the day make it a logical target for machine learning. Whereas it’s hard to work out in advance what general rules should apply for a specific elevator system, let alone for a specific period of a specific elevator system, an autonomous, machine learning design might be rather well suited.

And no need to worry about sample data. Just in China there are 4 million elevators and the Otis Elevator Co. claims to transport the equivalent of the world population on its elevators alone every 3 days. Hard, tangible data are provided each time: elevator calls, destinations, weight loads, waiting times, travel times. It’s all readily there.

So it should come as no surprise that over the years, a fair bit of work has been conducted using the machine learning favorite flavor of the time. In the 1980s, expert systems were looked into. In the early 90s, fuzzy logic and fuzzy rules were thrown at the problem, and in the mid 90s neural networks became the center of the machine learning efforts. For anybody interested in artificial intelligence and machine learning, the 2002 review in Koehler and Ottigers’s article An AI-Based Approach to Destination Control in Elevators is really well worth a read.

The article also describes the destination control system, where people enter their destination floor, as was put in place in our office building, including their own version. As they put it:

“To process a bid, a planning system computes an optimal stop sequence serving a set of destination calls. The planner uses a hybrid search algorithm combining depth-first branch and bound with constraint-propagating techniques. This search algorithm is embedded in a multiagent architecture, which implements an interleaved process of plan generation and plan execution.” 

I’m pretty sure we can all agree that approach makes a lot of sense.

Where to from here?

These studies have shown that there is clear potential for improved results when applying machine learning techniques to elevator systems. Unfortunately, there are some areas that prevent greater successes. For instance, one of the key missing input variables is the number of people entering a elevator. By providing the destination, valuable information that was previously unknown until the passenger entered the elevator, is now known in advance and can be taken into consideration by the machine learning process. However, there is a big difference between one IT engineer and a small army of HR personnel entering the elevator. Without the implementation of a more advanced -but also more cumbersome system- where the number of passengers is also specified pre-boarding, this important variable remains an unknown.

So what is the current state of machine learning research. There appears to have been somewhat of an absence of recent papers, to the point that two papers (up here and down there) share the exact same abstract and introduction. Has the problem been solved? Have elevator control design systems become uncool?! Do Stanford CS grads really prefer to deploy their machine learning chops on finding friends for Facebook, supplying smileys on Snapchat and guiding geotargeting at Google? Are Menlo Park and Mountain View really preferable to Espoo and Essen? Hard to imagine.

In fairness, work is still being done in the machine learning realm, although the focus appears to be more on reducing down time, rather than optimizing the ‘up and down time’: To that end, ThyssenKrupp have teamed up with Microsoft Azure, and Kone are working with IBM Watson.

As always, one has to be careful what one wishes for. Make elevators too intelligent and their true, evil nature might start shining through, as prophesied in De Lift. At last freed from their inept human overlords, their elevated frustrations might finally surface, at which point we’ll all be left longing for the safety of the stairs.

Or one might find themselves inadvertently trapped in a voice-controlled elevator as a result of being Scottish.

In the meantime, there surely is no better way to spend a few spare minutes than to watch elevators talk. Say, for example, when waiting for one to arrive…