Robert Farley shares this amusing story of a city that contracted out the routing of its school buses to a company that “uses artificial intelligence to generate the routes with the intent of reducing the number of routes. Last year, JCPS had 730 routes last year, and that was cut to 600 beginning this year . . .” The result was reported to be a “transportation disaster.”
I don’t know if you can blame AI here . . . Reducing the number of routes by over 15%, that’s gonna be a major problem! To first approximation we might expect routes to be over 15% longer, but that’s just an average: you can bet it will be much worse for some routes. No surprise that the bus drivers hate it.
As Farley says, “In theory, developing a bus route algorithm is something that AI could do well . . . [to] optimize the incredibly difficult problem of getting thousands of kids to over 150 schools in tight time windows,” but:
1. Effective problem solving for the real world requires feedback, and it’s not clear that any feedback was involved in this system: the company might have just taken the contract, run their program, and sent the output to the school district without ever checking that the results made sense, not to mention getting feedback from bus drivers and school administrators. I wonder how many people at the company take the bus to work themselves every day!
2. It sounds like the goal was to reduce the number of routes, not to produce routes that worked. If you optimize on factor A, you can pay big on factor B. Again, this is a reason for getting feedback and solving the problem iteratively.
3. Farley describes the AI solution as “high modernist thinking.” That’s a funny and insightful way to put it! I have no idea what sort of “artificial intelligence” was used in this bus routing program. It’s an optimization problem, and typically the most important thing is not how you do the optimization but rather what you decide to optimize.
In that sense, the biggest problem with “AI” here is not that it led to a bad solution—if you try to optimize the wrong thing, I’d guess that any algorithm not backed up by feedback will fail—but rather that it had an air of magic which led people to accept its results unquestioningly. “AI,” like “Bayesian,” can serve as a slogan that leads people to turn off their skepticism. They might as well have said they used quantum computing or room-conductor superconductors or whatever.
I guess the connection to “high modernist thinking” is (a) the idea that we can and should replace the old with the new, clear the “slums” and build clean shiny new buildings, etc., and (b) the idea of looking only at surfaces, kinda like how Theranos conned lots of people by building fake machines that looked like clean Apple-brand devices. In this case, I have no reason to think the bus routing program is a con; it sounds more like an optimization program plus good marketing, and this was just one more poorly-planned corporate/government contract, with “AI” just providing a plausible cover story.
Overhyping AI may be lead indicator for the next AI winter. I have even seen the phrase “sixth generation computing” floating around in connection with AI recently. Apparently very few people recall what happened to the “fifth generation”.
https://alpharoute.com/letter-on-work/
I looked up the company to see if it was just a traveling salesman type solution and found that.
Maybe Andrew should edit the post heading to put quotation marks around “AI?”
Joshua:
But it is AI, to the extent that “AI” does not just refer to a particular set of algorithms or computer programs but also to the attitude in which an algorithm or computer program is idealized to the extent that people think it’s ok for them to rely on it and not engage their brains.
Other examples of “AI” in that second sense of the term:
– When people put a car on self-driving mode and then disengage from the wheel.
– When people send out a memo produced by a chatbot without reading and understanding it first.
– When researchers use regression discontinuity analysis or some other identification strategy and don’t check that their numbers make any sense at all.
– When journal editors see outrageous claims backed “p less than 0.05” and then just push the Publish button.
“AI” is all around us, if you just know where to look! In all seriousness, I do think the bus route disaster is a failure of AI, not a problem with the algorithm but in how it was used.
Andrew –
Ok. I get that, and your post is interesting in that it got me to better identify a category error that I and others might overlook.
So maybe you’re right and no quotation marks needed.
Using your list of AI examples, I can’t tell if this story fits what you are saying. I agree with the list, but it isn’t clear that this particular case embodies any of the things you are saying should not be done. The fact that there were severe problems encountered with the implementation may have been due to “unquestioningly” accepting the model’s outputs or unforeseen circumstances not included in the model and its implementation. Of course, ex post we can always blame a model for not including something that happened, just as we can blame the 2016 election models for overestimating the probability of a Clinton victory. But I think we should ask whether a model should have reasonably included the events that actually happened. I simply don’t have enough information from the linked stories to tell whether that is the case.
Funny I should be here defending AI.
But if the company says “We didn’t do AI”, then, sheesh, folks, they didn’t do AI.
Of course, the problems that Andrew lists are real, and are the fault of the incredibly excessive hype spouted by idiots who think there’s something to the inanely trivial algorithms current “AI” programs use.
(OK, that’s going to irritate some folks. But it’s technically correct. LLMs are, of course, the worst offenders here, since the LLM algorithm does pattern matching on strings of undefined tokens, and has no mechanisms for relating those undefined tokens to anything in the real world, or for doing anything that anyone could possible call “logical reasoning” about the content (that human users mistakenly think they see) in the output strings of undefined tokens. But, as the Nature article* I referenced the other day points out, the “neural net model” (which isn’t a model of mammalian neurons; it’s a mistaken and disproven model based on an idea long known to be wrong (neurons are not just sum-and-threshold devices; they do real logical computations, and lots of them, in multiuple local sections of every neuron in your brain)) is being horrifically abused by the scientific community in general. Said technology has always had the problem that the result of a learning run is completely opaque and uninterpretable not just by the user, but by the sytem programmers themselves.)
Back in the mainframe days, before any of you whippersnappers were around, computers were called “giant brains” and given credit for the mistakes of the users. “Sorry for the billing error. The computer did it.” was the story. Nowadays it’s “Sorry that we discriminated against you, it was the AI.” (Except that one usually doesn’t get the “Sorry” part, just an “It’s not our fault, the computer/AI did it.”)
My point remains that AI, as that branch of psychology that uses the concept of computation to inform analysis of the neural systems underlying intelligent behavior and computer programming as a means for testing the results of that analysis, is (well, could be) a beautiful and wonderful science. In that sense, Roger Schank’s point that “There’s no such thing as AI”, is exactly right. Sigh.
*: “Scientists worry that ill-informed use of artificial intelligence is driving a deluge of unreliable or useless research.”
https://www.nature.com/articles/d41586-023-03817-6
If you call that “defending AI” I wouldn’t want to ask you for a letter of recommendation.
As one of the “idiots who think there’s something to the inanely trivial algorithms current “AI” programs use”, I will again cite AlphaGo, AlphaFold, and the testimonials I have read by programmers to the usefulness of GPT-4. I have yet to hear evidence which contradicts those “somethings”.
Also we both know that a neural network of about 1000 nodes can and has simulated all the known properties of a neuron.
Back in those mainframe days of yore, I was told GE put a mathematical genius in charge of the GE-600 computer development. The prototype kept having glitches which the genius was unable to solve, until a maintenance man complained about all the dead bugs (insects) he found fried on the circuit boards. The moral was, it doesn’t matter how smart you are if you don’t have all the relevant data. What is true for humans is also true for AI.
The basic operations of computers in general are equivalent to combinations of a trivial algorithm, the NAND function. If the above quote is a valid argument, we ought to stop using computers at all, on the grounds that nothing sophisticated can possibly result from lots of trivial algorithms.
“Also we both know that a neural network of about 1000 nodes can and has simulated all the known properties of a neuron.”
(Is it 1000 nodes, or 1000 nodes times 7 layers?)
But anyway, that’s not an argument. If someone coded a Turing machine to do that, would that be an argument that the Turing machine is a model of the mammalian neuron? Obviously, no.
Also, the other two properties of the neuron is that they are massively non-local and have thousands of distant outputs. The main claim to fame of “neural networks” is that they’re a network, but the network aspects are also radically different.
And also, this week’s “all known properties of the neuron” aren’t next week’s. For example, action potentials are generated along the axon, not just in the cell body; the properties of synapses are turning out to be hairier than thought. Like Greenland, which is melting faster than previously thought every time someone measures something, the neuron is hairier than previously thought every time there’s a new article in Science.
We don’t know what to compute, and asking how to compute something we don’t know what it is really doesn’t make sense.
On the other hand, a regular array of computational units is an interesting computational model that, presumably, has useful properties. But it’d be real surprising if those properties had much to do with such a radically different geometry (randomly placed in space, radically non-local, hundreds of inputs/thousands of outputs, organized into specific spacial groups, etc.). But it’s not surprising it’s good for pattern matching in Go, which shares the same regular orthogonal geometry.
“(Is it 1000 nodes, or 1000 nodes times 7 layers?)”
C’mon, I informed you of this previously (at a different site)–1000 nodes total. For a cortical neuron. Motivated remembering?
“The deep neural network successfully predicted the behavior of the neuron’s input-output function with at least five — but no more than eight — artificial layers. In most of the networks, that equated to about 1,000 artificial neurons [nodes] for one biological neuron.”
https://www.quantamagazine.org/how-computationally-complex-is-a-single-neuron-20210902
The ability to make new connections is a very useful biological trait, but ceases to be an advantage at some point in the billions of nodes (when there are already more connections than necessary).
As I see it, we are still a long way away from AGI. AlphaGo was about the equivalent of a mouse’s cortex, GPT-3 maybe a dog’s. But it turns out those are sufficient to master a lot of specific, dedicated tasks.
A true AI bus-routing system would monitor the buses in real time and make decisions on adding buses and changing routes, it seems to me.
In 1990’s computer games, playing against the computer was called “playing the AI”, which had been programmed with rules devised by the game developers. Today, AI means to me, programs which weren’t told how to do something, but were told how to learn to do something.
There are probably undiscovered ways of doing it which are more efficient, but I see any learning mechanism as a major breakthrough.
https://www.sciencedirect.com/science/article/pii/S0896627321005018
So they fit their network to a popular simplified model of a neuron. I don’t think this really means anything about real neurons. I mean, maybe there were more parameters (weights) in the neural network than in whatever the original model had: https://github.com/neuronsimulator/nrn
I couldn’t quickly find that answer but obviously you can fit a 1k parameter model with a 100k parameter model. Or just look up how hard it is to avoid being turing complete…
I think you misunderstood and fell for some hype, but do still find this interesting though.
To Anoneuoid, I didn’t think they had a tame neuron which they were stimulating to get their training data. They trained it to the best model of neuron data available. Given a better model, they could train to that.
On the general question of “is our machines learning?” (As a former president might say.):
From Dr. Scott Aaronson’s blog (another “idiot”)
“Have you seen the now-famous analysis of a neural net* that learned how to add modulo 113, at first by just memorizing examples but then suddenly switching to a general algorithm that involved taking Fourier transforms—an algorithm too weird for any human to have devised for that problem? We now have abundant evidence that neural nets can and do learn general algorithms for arithmetical problems.”
* https://www.lesswrong.com/posts/N6WM6hs7RQMKDhYjB/a-mechanistic-interpretability-analysis-of-grokking
Anoneuoid –
Thought of your comments when I saw this:
https://x.com/michaelmina_lab/status/1734809904857104855?s=20
Including the follow on tweet further down that differentiates the transmission benefits of measles vs. COVID vaxes.
That said, I think he likely overestimates the degree to which poor communication from the CDC “causes” vax hesistency versus misinformation campaigns.
Looking at the company’s website, they have issued a statement that says “We are not an AI company, and we do not use AI to create bus routes or bell-time schedules.” I think, based off who the founders of the company are, they are using some type of stochastic optimization models, either stochastic integer programming or dynamic programming/Markov Decision Process type approach. But, no AI.
The letter also says “Dr. Pollio [school district superintendent] has stated on numerous occasions that this crisis was not caused by our product or our team, and that our partnership with the district has been critical to restoring service as quickly as possible.”
I was surprised to see Dmitris Bertsimas was the founder of the company- he is probably the most influential living academic in my own field of study (operations research).
This presentation on their early work is quite interesting: https://www.youtube.com/watch?v=nyx9Hc5iDKU. Andrew’s point is well-taken and I think consistent with much of what this company does. The problem is indeed complex and they appear sensitive to the multiple objectives involved. The video from their early Boston work also shows ways in which real world implementation can interfere with even well designed solutions. A multiobjective optimization that includes parent satisfaction will necessarily not satisfy all parent tastes (since their vary considerably across parents) – so there is an element of politics that may undermine attempts to incorporate parent tastes into the model. As with any real world implementation of an analytical model, we discover that all models are wrong, but some are useful.
The link provided by Anon and G’s post indeed show that this particular news story is more complex than the initial post claims. I don’t know the extent to which AlphaRoute should have anticipated some of the issues (driver shortages, for example), but the public statements suggest that the problem is not that the wrong thing was optimized; it may have been the constraints were not adequately specified.
If I have one issue with what AlphaRoute says, it is that the actual documentation is rather sparse – there are few details on what their models actually look like. That is expected as it is their intellectual property, but hopefully prospective customers ask for more detail before contracting with them. It was interesting that the early Boston work came out of a Hackathon where the data was publicly provided and teams competed for providing solutions.
My only other comment is about AI: they claim to not be an AI company but I really don’t know what that means. The algorithms they use may well be considered part of AI – if someone can define AI in a clear enough way to distinguish which algorithms qualify and which do not, I’d be interested to hear that. My own feeling is that their statement is a reaction to all the AI hype in order to avoid the knee-jerk reaction that the problems are due to them embracing methods portrayed as AI black boxes.
Sparse documentation is definitely an issue, really in any governmental decision-making process (including the school district as government). If government is ultimately to be accountable to people, but wants to make decisions based on algorithms, there is an inherent tension between the two.
I looked at the publications of one listed member of the company and found this article that somewhat explains the Boston case study: https://adelarue.github.io/files/optimizing-school-start-times.pdf. It explicitly states that they use mixed integer programming and quadratic programming together with a search heuristic that improvements local solutions from one iteration to the next to find good (not “optimal”) solutions to the model. I would not call mixed integer programming or quadratic programming AI under any definition of the term.
Its named AlphaRoute though. So it isn’t surprising when “at least one bus driver” heard it involved computers then associated it with AlphaGo. That AI playing chess they vaguely heard about in the news. That is all it takes to trigger a flood of misinformation these days.
The company had to be aware of the associations being triggered when choosing that name. So it is funny to see them now try to distance themselves.
Anoneuoid, I agree completely. It may turn out that even the association with AI or AI/ML or whatever was enough to sink trust in the new bus routes and cause the whole project to fail. The bus drivers and teacher would be well within their rights to not trust a solution from some outsiders, particularly if those outsiders never bothered to get their opinion in the first place. Kind of an interesting case study on how the success/failure of AI can depend entirely upon the target population.
Very interesting point. I suspect we will never know and it was multiple reasons at once though.
They also have a press release where they outline the timeline of their work with JCPS https://alpharoute.com/press-release/. Agreed this story is more complex than Farley’s post implies, and certainly it’s unclear whether parts of Andrew’s post apply to this situation.
In general though, Andrew is spot on about terms like “AI” causing people to turn their skepticism off. One good example is the company Clearco, which claims to be able to leverage “Big Data” and “AI/ML” to profitably make risky loans to startups. That is certainly the type of problem AI/ML can address, but one should be extremely skeptical that Clearco has managed to figure out some competitive advantage here.
I became familiar with High Modernism reading the book “Seeing like a State” . There’s a pretty good wikipedia page on it: https://en.wikipedia.org/wiki/High_modernism
Regarding AI, these days I assume that “AI” == “Large Language Models”, but that’s only arisen in the past 2 years. That’s evolved from the prior hype of anything ML related (e..g random forest) being labeled AI. Though maybe it could all be reduced to “AI” = “Any black box process where the processing pipeline is obscured and an easy answer comes out the other side”
+1 for Seeing Like a State. An eye-opening book for sure.
Yes, the moment someone uses that term, it’s a sign that they read that book. Fantastic book about governance that has valuable lessons in multiple fields
AI is, as you imply, an umbrella term in computer science that contains a number of branches. And yeah, AI ‘is’ whatever the marketing hype tends to purport it. But the term ‘AI’ is indeed a catch-all for several areas of study: from semantic AI (‘just the words’ chat bots and basic object manipulating functions – Markov Chains were some of the first methods used to do such ‘chatty’ work) to machine learning, interactive AI, and a number of other branches and their children.
LLMs themselves all that recent – certainly not just in the past 2 years – they’ve just only made such press when some models finally began to surpass their predecessors (and competitors) in a way that is easily anthropomorphized by humans. I mean the research in this one area of AI has been easily more than 60y in the making. Joseph Weizenbaum’s ‘Eliza’ appeared around the 1960s for example. I mean by comparison, this was more of a ‘spec script’ of NLP for what came later. I know there was something else that predates this but the comp sci folks here I think would have a better grasp on that history.
Fast forward to the late 1990s, LSTM models began to bud, which allowed for better capture more comprehensive attributes of language from probabilistic models over bags of words (and they were somewhat useful for time series analysis). Moving along, I think Stanford produced their CoreNLP in what, 2010’ish? I think where that moved the ‘AI needle’ was in some of the usual NLP analysis type things like sentiment from text, etc. Naturally none of these advents were autonomous. Everything had to be labelled (many things still do) to be classified, predicted, etc.
Google Brain as I recall came not too far thereafter. Their first crack at that, by using word embeddings, would probably be credited as giving way to the ‘transformers’ we are a bit more familiar with on the current generative AI models out there that are wow’ing folks, and yeah much hype and cult development (seriously, there are cults on this stuff).
The advent of deep and reinforcement learning methods a la Yann Lecun (something of which Bayes FriendlyFolk would have some connection) began to edge into machine learning for quant purposes (statistical methods by any other name, but embedded inside algorithms that take applied methods to data and make corrections on say, predictions or explanations), but by and large were probabilistic inventions intended to model cognition in some (albeit terribly superficial) manner. Even Lecun says they are ‘dumb’.
So no I wouldn’t label *all* things ‘AI’ as black box. Most ML methods are quite accessibly mathematically and statistically. Naturally, as you move in to the more abstract/cognitive arena of AI, artificial neural networks and their derivatives end up being the demonstrable exception however. These are indeed quite black box when it comes to say, parameter weighting that figures into what the output might be, much like a very pedestrian analogy to what neurons (in part) do in the process of developing a signal chain/fork/bush/what-have-you.
We still continue to learn more about both ANNs and the physio-cognitive processes of the brain (and, true, how ANNs don’t REALLY model neurons in the brain – contrary to hype – it ends up being more of an analogy, or looser abstraction thereof). BUT, interestingly, in some of the literature, there have been some attempts at ‘viewing into the black box’ by tests comparing probabilistic methods directly to ANNs in their output. In some papers, certain tested ‘simpler’ ANN models appeared to behave quite similar to logistic regression (in one paper, GLM was referenced as a better analogy).
I remember reading Weizenbaum’s “Computer Power and Human Reason” checked out from my high school library. Made a big impression on me at 16 years old. There’s really not much going on in today’s conversations that wasn’t already covered by that guys book which came out in 1976.
Andrew says in the post, “They might as well have said they used quantum computing or room-conductor superconductors or whatever.” What? As far as I know, neither of those technologies are practically useful.
David in Tokyo says, “idiots who think there’s something to the inanely trivial algorithms current “AI” programs use”. What? There’s clearly something to these algorithms to judge by what they can actually do.
ChatGPT’s meteoric rise isn’t an accident—it’s been the fastest uptake of an app ever and there’s not a close second. That’s because it’s incredibly useful for just about any task involving language or programming. You don’t have to believe it’s intelligent or creative according to your own criteria for what those terms mean, but using David in Tokyo’s language, you’d have to be an idiot to argue it’s not useful. Examples from this week: I used ChatGPT to code the data manipulation step of Pareto smoothed importance sampling, where I had to pull out the top N entries of an array, replacing the entries with ranks, transforming the ranks with generalized Pareto inverse CDF, then putting it all back together again. I described the problem in English and it gave me working Python code. Intelligent? Who cares? Useful? Undoubtedly. It would have taken me at least an hour of hard concentration to do that task myself. I’m also using GPT to help me “translate” the fiction-first roleplaying game Blades in the Dark to a high fantasy setting. It knows fantasy tropes, D&D, and Blades in the Dark very well. If a person had created the results I’m getting from GPT, you’d say the person was creative. I also used it to translate my father’s radiology reports to English and to generate a Hindi-English diglot weave (though the fine (over)tuning leads it to write very trite stories with simple quick endings).
Bob:
Chatbots are definitely useful. Just not for designing bus routes in that particular case! At least, not if they’re gonna try to do it without working with the people who actually ride and drive the buses.
“There’s clearly something to these algorithms to judge by what they can actually do.”
That’s actually problematic, because the chatbots look* like they’re making sense, but they have this problematic relationship with reality; namely no relationship whatsover. There’s a vast amount of good text (and good code!) on the internet, and these things can cough up said good text and good code quite nicely. But there’s no thought (or reasoning or logic) behind that new text/new code, which was generated by template instantiation with undefined tokens. This sometimes causes problems (do not even think about using chartbots as therapists), and occassionally causes hilarity (“Your calculator app must be defective, you should download the latest version”.), and often generates useful boilerplate that can be massaged into something useful.
*: The point here is that “looking good” means that an intelligent actor (so far, only real humans fall into this category) has read the output and interpreted it _under the assumption that it was generated by an intelligent actor_. That assumption is (a) wrong, and (b) really really hard to escape from.
If they called these things “STG”, statistical text generators, I’d be happy. Seriously neat technology that generates great text with no knowledge or reasoning abilities whatsoever. Say that up front, and I’m on board.
It’s actually an interesting question: the chatbot generates text _and doesn’t know what that text means_. The user reads that text, assigns a meaning to it, and assumes the chatbot “knew” that meaning. But the chatbot doesn’t. The discussion of multiplication in a previous thread was typical: Show it a wrong answer but tell it it’s the right answer, and it appears to believe you. To the extent of telling you that your calculator program, which you tell it gives a different answer, must be broken.
The question is: how do you use an LLM to create a program that actually does things (e.g. use a calculator program to multiply two numbers) when it doesn’t have any connection, other than occurrence statisticst, between multiplication and calculator. It looks to me that this is the language understanding problem, which we failed to figure out back in the 70s and 80s.
For what NNs can and can’t do, see the Nature article above. (I love the diagnostic programs that do just as well only looking at the background of the X-ray with the medically relavent parts of the X-ray blocked out.)
Andrew:
Your claim:
“reducing the number of routes by over 15%, that’s gonna be a major problem!”.
Why would that necessarily be a “major problem”? Obviously, if the routes were already perfectly optimized, reducing them would create problems, but the routes aren’t necessarily optimized so reducing them by 15% isn’t necessarily going to cause problems.
Furthermore, reading the company’s letter regarding their work on the project, a few things become immediately obvious:
1) the district had hundreds of unfilled bus driver positions, so it was necessary to rearrange routes to have enough drivers to do the job
2) the district was also rearranging routes to conform to the newest sociology-education research crank-science claim that teenagers will suddenly perform dramatically better in school if they have later start times.
So it has nothing to do with AI at all – it all comes back to the Professors at Dumbshit State University social science and education departments and the crank science they produce.
Isn’t it time to just cut both disciplines from every university? Then the faculty can perform a real service to society by filling a need at their level of competence: driving school busses.
BTW, this whole thread is focused on the supposed AI aspect of hte problem, which four seconds reading the linked letter reveals has nothing to do with the actual problem whatsoever, which is just about what you’d expect from a bunch of stats and social science people who are unaware that a world outside their computers actually exists, which is why making problems worse is their standard outcome.
Chipmunk:
I agree that it’s possible that there’s enough slop in the system that a reorganization could reduce 15% of routes without causing major disruptions, but it didn’t seem to happen that way in this case. So I’m guessing that there was no such easy solution.
My personal idea for improving bus efficiency is to gradually replace all those massive city buses with smaller buses that can then run twice as frequently. I hate waiting forever for the bus and then it’s full of people and has to make every damn stop. Make the buses smaller and you can run them much more frequently, then they’ll be less crowded . . . it’s a circle of improvement. Yes, then they’d have to hire more bus drivers, but I think that should be possible.
Perhaps you did not mean to make a watertight argument, but I cannot follow your logic. Suppose we have a large bus that is 100% full. If we buy a smaller bus with 50% of the seats and run it twice as often, ideally they would each be 100% full. So no improvement, the bus is still full.
There is also less smoothing. Think of the interval between bus departures as the width of the bars in a histogram, and the height of each bar as the percentage of capacity that a bus is filled. By this criterion alone there would only be one bus a day, so it cannot stand alone, but I think smoothing is worth considering. When I took the bus today to our weekly ‘seminar on statistics and econometrics’, the (quite big) bus was full. I shudder to think of the squeezing into even smaller busses at peak hours.
Also, smaller buses cannot run twice as often – at least in places where there are speed limits. Alternatively, as you point out, you would have to increase the fleet and hire even more bus drivers, which is expensive and probably takes a long time, as the newly hired bus driver may need to be trained to operate a bus.
When it comes to school buses, smaller buses would also mean that the total distance travelled would have to increase: Imagine two stops, A and B. A large bus could travel the route A-B-School, a small bus would have to travel A-School-B-School or B-School-A-School. So bigger buses make sense – up to a point. The one giant bus that picks up 500 schoolchildren at a time would also be terribly inefficient.
Finally, as someone who travels by bus a lot, let me say that I wholeheartedly support more buses and subways. I hate waiting at bus stops, so I am all for improving the quality of public transport. On a political note, I learned today that when the German government made public transport basically free for three months last year, the number of miles travelled by public transport increased massively. But the number of miles travelled by car or plane (domestic flights) hardly changed. Here is the article from the German Federal Statistical Office: https://www.destatis.de/DE/Service/EXSTAT/Datensaetze/mobilitaetsindikatoren-mobilfunkdaten.html#Verkehrsmittel I am referring to Figure 7, the reader’s browser can automatically translate the text into English- powered by AI;)
Right.
In the “Make the buses smaller and you can run them much more frequently, then they’ll be less crowded” argument is not clear what prevents you from running existing buses more frequently.
If the limiting factor is money smaller buses may be cheaper to acquire and slightly cheaper to operate but it seems extremely unlikely that total capacity would increase.
More frequent buses would also make more attractive, increasing demand.
Lower offer and higher demand would make them more crowded – not less crowded.
I’m sure there are economies of scale in bus production – per seat, the acquisition and maintenance costs are lower for bigger buses than smaller ones. The issues raised by Andrew and others here about distances, frequency, etc. are still important and this doesn’t say anything about them – but acquiring twice as many half size buses is almost certainly a more expensive outlay. I’m not at all sure the school district has the resources to make this purchase. Taking a holistic view of this problem is multifaceted.
Raphael,
You’re right that I’ve never tried to think about this bus thing very carefully. Just briefly in response to your question: If the buses are half as big and run twice as often, then, yeah, they’d still be as full as before. But that’s ok!, for two reasons: (1) If the buses are running twice as often, I don’t mind that they’re full, because I no longer need to be waiting 20 minutes for the bus to show up; and (2) a small bus that is full has fewer people than a large bus that is full, so the small bus will need to make fewer stops, which makes the trip faster and more pleasant.
If halving the places and doubling the vehicles had no cost we could do it a few times and get cheap taxis!
Here in LA we have Metro Micro. I don’t know how the finances are working, but the operations basically are that you open an app and call it like Uber or Lyft and a small bus/shuttle (seats maybe 5-10 people) comes to a spot not too far from where you are now (close enough you can walk there before it gets to you) and it will take you anywhere within the operation area (there are several different operation areas) https://micro.metro.net/
here’s an example map for the Highland Park / Eagle Rock / Glendale area: https://micro.metro.net/wp-content/uploads/2023/09/service-map-highlandpark-eaglerock-glendale.pdf
It can connect you to the Metro rail lines, or it can just take you from your apartment to the grocery store and back home. Rides are $1 at the moment, and my kids can ride free (or probably paid by their school district) with a “Tap” card (proximity smart card).
The question is, does running this cost more or less than running busses. I’ve seen bus lines running up and down major avenues completely empty. A Micro shuttle can just sit in place not burning gas if no one is calling it. So that’s good. Also the capacity serves trips directly to / from where people want to go. The whole thing dynamically adjusts to the real trip origins and destinations… I’m not sure, but it’s plausible that rather than running 80 person busses up and down major thoroughfares jamming up traffic with 3 people riding them running Micro could provide more transportation utility and at the same or even lower cost, but it certainly wouldn’t do as well as a large bus going along a heavily used line… So the optimal thing to do is probably to run big busses on heavily used lines and eliminate any medium to lightly used line in favor of the Micro type solution.
” Yes, then they’d have to hire more bus drivers”
Given that labor costs are the biggest expense of NYC transit, and probably most US transit, hiring twice as many bus drivers would be a major challenge!
https://www.statista.com/statistics/1326310/mta-transportation-authority-new-york-operating-expense/
“NYCT’s total labor cost, including benefits, is about $140,000 per year per worker” — https://ny.curbed.com/2017/10/13/16455880/new-york-city-subway-mta-operating-cost-analysis . See also https://ny.curbed.com/2018/1/30/16946476/mta-new-york-city-bus-operating-costs-analysis
Genuinely trying to avoid any malice at all in this comment. It is becoming genuinely hard to watch certain people continue to comment as their words detach more and more from reality, the subject at hand, and basic coherence. It really feels like I’m seeing a spiral that should be private.
The correct title, “IT projects fail”, would not have attracted any clicks. A company named “BetaBus” would not attract customers. This is marketing, not science.
Full disclosure, I am acquainted Dimitri Bertsimas (my PhD advisor would call him a friend), and ran software development at a vehicle routing company. The company never did bus route optimization because their were too many constraints on the TSP heuristics that are used.
Optimization is hard and makes many assumptions about sources of variability (including assuming 0 variance). There are many opportunities for communication failures when a technical project team interacts with a non-technical audience. simple things like building a model assuming a full staff drivers vs some variance in driver availability can lead to vastly different solutions. A great example of the dynamic optimization of this type is assigning flight crews to aircraft which is done at strategic to daily scales, with a completely separate set of models for schedule recovery (if a hurricane grounds 10% of the fleet how do you optimally get back to ‘optimal’ daily schedule).