00:00:01
All right, so here's something crazy.
00:00:03
Turns out most AIs are horrible at real
00:00:06
life decision-making. I'm talking about
00:00:08
the stuff that actually matters.
00:00:10
Planning routes, hitting deadlines,
00:00:12
keeping deliveries on time without
00:00:13
wasting fuel. We always hear that AI is
00:00:16
changing the world, right? But ask it to
00:00:18
schedule 50 drivers or plan a delivery
00:00:20
run across a city and it breaks down
00:00:22
like a cheap GPS. That is until now.
00:00:26
Because what Deep Mind just pulled off
00:00:28
might actually fix one of AI's biggest
00:00:31
blind spots. And if it works at scale,
00:00:34
it could change how your groceries get
00:00:36
to your door, how your doctor schedules
00:00:38
surgeries, even how cities avoid traffic
00:00:40
jams. It's like giving AI actual street
00:00:43
smarts. And yeah, it's a big deal. So,
00:00:47
let's talk about it. Okay, so here's the
00:00:49
deal. Planning delivery routes,
00:00:51
scheduling workers, or even figuring out
00:00:52
the best way to get supplies to a store
00:00:54
is crazy hard. These are what tech folks
00:00:57
call combinatorial problems, which just
00:01:00
means there are a ton of possible
00:01:02
choices and you need to pick the best
00:01:04
one without breaking the rules. Like
00:01:07
making sure a delivery truck doesn't run
00:01:09
out of space or miss a time window. For
00:01:12
example, let's say a delivery driver has
00:01:13
to hit 50 houses in a city, each with a
00:01:16
specific time they need their package,
00:01:18
and the truck can only carry so much.
00:01:20
Finding the perfect route is like
00:01:21
solving a puzzle with a billion pieces
00:01:23
and doing it fast. Good luck. In the
00:01:26
tech world, they call these problems NP
00:01:29
hard, which basically means they're so
00:01:32
complex that even the fastest computers
00:01:34
would take forever to find the perfect
00:01:35
answer. Now, you might think, "Hey,
00:01:37
we've got AI, right? Can it just figure
00:01:39
this out?" Well, here's where it gets
00:01:41
tricky. AI, specifically neural
00:01:43
networks, are excellent at spotting
00:01:45
patterns like recognizing faces in
00:01:47
photos or predicting what movie you
00:01:49
might like based on your history. They
00:01:51
work best with smooth continuous data
00:01:53
like images, speech, or text. But when
00:01:56
it comes to rigid, all orno decisions
00:01:58
like planning a truck route with strict
00:02:00
timing and capacity rules, they fall
00:02:03
short. These kinds of tasks need
00:02:05
clear-cut logic and constraint handling,
00:02:07
which neural networks just aren't built
00:02:09
for. On the other hand, traditional
00:02:11
methods that can handle these decisions,
00:02:13
like advanced mathematical solvers, are
00:02:16
incredibly slow and demand perfect
00:02:18
information, which isn't realistic when
00:02:20
you're trying to move a mountain of
00:02:22
packages in real time. So, these
00:02:24
researchers were like, we need a way to
00:02:26
make AI better at these tough logistical
00:02:29
puzzles. And they came up with something
00:02:31
awesome, a way to teach AI to think like
00:02:34
a human planner, but way faster. They
00:02:37
created a trick called MCMMC
00:02:39
layers. Don't let the name scare you.
00:02:41
It's just another way of saying they
00:02:43
gave AI a tool to explore different
00:02:45
options and pick the best one. Kind of
00:02:48
like how you'd try different routes on
00:02:50
Google Maps to avoid traffic. By the
00:02:51
way, MC MC stands for Marov Chain Monte
00:02:54
Carlo, but let's just call it a smart
00:02:56
explorer. It's like a GPS that doesn't
00:02:59
need a perfect map. It checks out nearby
00:03:01
options like swapping two stops or
00:03:03
rerouting a truck and decides if it's a
00:03:06
good move based on some clever rules.
00:03:09
Here's how it works in a way we can all
00:03:11
get. Imagine you're planning a delivery
00:03:13
route and you've got a decent plan, but
00:03:15
it's not perfect. The smart explorer
00:03:17
looks at small tweaks like moving one
00:03:20
stop earlier or swapping two houses and
00:03:23
checks if they save time or gas. It uses
00:03:26
a method inspired by something called
00:03:28
simulated annealing, which is like
00:03:31
slowly cooling a hot piece of metal to
00:03:34
make it stronger. In this case, it's
00:03:36
cooling the AI's choices to focus on the
00:03:38
best ones over time. The researchers
00:03:41
turn this into a layer that fits right
00:03:42
into a neural network, so the AI can
00:03:45
learn from its mistakes and get better
00:03:47
at picking routes, all while staying
00:03:49
fast and flexible. What's really cool is
00:03:52
that this doesn't need a perfect
00:03:53
solution to work. Old school methods
00:03:55
relied on something called exact
00:03:57
solvers, which are like trying to solve
00:03:59
that billionpiece puzzle by checking
00:04:01
every single piece. That takes forever,
00:04:03
especially for big problems. This new
00:04:04
approach uses what's called local search
00:04:07
huristics. Think of them as quick, smart
00:04:10
guesses that get you close to a great
00:04:12
solution without obsessing over
00:04:14
perfection. The researchers made sure
00:04:16
these guesses are differentiable, which
00:04:18
just means the AI can learn from them,
00:04:20
like how you learn from trial and error
00:04:22
when planning a party. They also used
00:04:25
something called fential young losses.
00:04:27
Okay, I know another techy term, but
00:04:29
stick with me. It's like a scorecard
00:04:31
that tells the AI how close its plan is
00:04:34
to the best possible plan. Even if the
00:04:36
AI only tries one quick guess, this
00:04:38
scorecard keeps the learning process on
00:04:40
track, which is huge because it means
00:04:43
faster training and less computing
00:04:44
power, they tested different ways to
00:04:46
start the AI's guesses, like starting
00:04:48
with a known good plan called ground
00:04:51
truth, or a slightly improved one using
00:04:54
a huristic, which is like giving the AI
00:04:56
a head start with a rough draft of the
00:04:58
route. Now, let's get to the juicy part,
00:05:00
the results. They tested this on a super
00:05:03
tough problem called the dynamic vehicle
00:05:05
routing problem with time windows or
00:05:09
DVRPTW from a big tech competition
00:05:11
called Euro meets NUR IPS 2022. Picture
00:05:16
a city where delivery requests keep
00:05:18
popping up all day and you've got to
00:05:20
assign routes to trucks while making
00:05:22
sure they hit each stop on time and
00:05:25
don't overload the trucks. It's like
00:05:27
playing a real-time strategy game where
00:05:29
new challenges keep coming and you've
00:05:30
got seconds to decide. The researchers
00:05:32
use a setup where each truck's route is
00:05:35
planned in waves with new requests added
00:05:38
and old ones cleared out as you go.
00:05:41
Their AI powered by these MCMC layers
00:05:44
was a rockar. When they gave it just one
00:05:46
millisecond to decide a blink of an eye,
00:05:49
it came up with routes that were only
00:05:50
7.8% 8% worse than a perfect plan that
00:05:53
knows all future requests, called the
00:05:56
anticipative baseline. Compare that to
00:05:58
the old method, which used something
00:06:00
called perturbation, basically adding
00:06:02
random noise to shake things up, and it
00:06:04
was a whopping
00:06:06
65.2% worse. That's like the difference
00:06:08
between getting your package in an hour
00:06:10
versus next week. Even when they gave it
00:06:13
more time, like 1,000 milliseconds,
00:06:15
their method hit 5.9% compared to 5.5%
00:06:19
for the old approach, showing it's
00:06:21
neckand-neck with the best, but way more
00:06:23
practical for real world use. They also
00:06:26
found that starting the AI with a good
00:06:29
plan, like the perfect routes from the
00:06:31
competition's baseline or a slightly
00:06:32
tweaked version, made it perform even
00:06:35
better. For example, at 100
00:06:36
milliseconds, the tweaked start got them
00:06:38
to 5.9% relative cost. super close to
00:06:42
ideal. They played with a setting called
00:06:45
temperature, which is like telling the
00:06:46
AI how much to explore new ideas versus
00:06:50
sticking to what's working. A
00:06:52
temperature of 100 was the sweet spot
00:06:54
when starting with a good plan, but
00:06:56
lower temperatures helped when starting
00:06:58
from scratch. To make sure this wasn't
00:07:00
just a fluke, they tested it on simpler
00:07:02
problems, like picking the best
00:07:04
combination of items. They found that
00:07:06
their smart explorer got super close to
00:07:08
the perfect answer, even with just a few
00:07:10
tries. Running multiple explorers at
00:07:12
once was like having a team of planners
00:07:14
working together, speeding things up
00:07:16
without losing accuracy. They also
00:07:18
proved that their method is rock solid
00:07:20
mathematically, using terms like
00:07:23
convergence guarantees to show that it
00:07:25
reliably learns the best solutions over
00:07:27
time, even for super complex problems.
00:07:30
In the vehicle routing tests, they used
00:07:32
specific tweaks like swapping two stops,
00:07:36
moving a delivery to a different spot,
00:07:38
or flipping part of a route, and made
00:07:40
sure they followed the rules like not
00:07:43
overloading trucks or missing time
00:07:45
windows. They ran these tests on a
00:07:48
single CPU using 30 problem sets for
00:07:51
training and 25 for testing with up to
00:07:55
100 requests per wave. The results were
00:07:58
averaged over 50 runs to make sure they
00:08:00
weren't just getting lucky. And they
00:08:01
used a graph neural network, like a
00:08:04
super organized spreadsheet, to handle
00:08:06
the data. So, why should you care? This
00:08:09
tech could make deliveries faster and
00:08:11
cheaper, which means lower prices for
00:08:13
you and less stress for companies.
00:08:15
Imagine your food arriving hot, your
00:08:17
packages showing up on time, or even
00:08:19
hospitals scheduling surgeries more
00:08:21
efficiently. It's not just about trucks.
00:08:23
It could help with anything that needs
00:08:25
smart planning, like organizing events
00:08:27
or managing traffic. The catch, it's not
00:08:30
perfect yet. The researchers had to
00:08:32
tinker with the AI's inner workings to
00:08:34
make it work, which isn't always easy.
00:08:37
But they're already thinking about ways
00:08:38
to make it even better, like using
00:08:40
smarter shortcuts to explore bigger
00:08:43
options. All right. Now, is this how it
00:08:45
starts? AI solving delivery routes
00:08:48
today, deciding who gets health care
00:08:51
tomorrow? Hm. I want to know what you
00:08:53
think. And hey, if you liked this
00:08:54
breakdown, hit that like button,
00:08:56
subscribe for more Wild Tech stories,
00:08:59
and thanks for watching. Catch you in
00:09:01
the next one.
