The algorithms being used in machine learning are not actually new; they're decades old, and many of them were first used in problems in systems engineering. As early as the 1990s, researchers realized that the field of AI was studying the same concepts with different terminology, but for a variety of factors it was the AI space that found the most success. Nevertheless, we're coming back full circle as we see the integration of data, software, and physical equipment starting to blend together. What comes next in the world of data? And how can we learn from the technology of the past?
3. Which came first?
Take a moment to mentally order these chronologically
A. David Hasselhoff was born
B. Rosalind Franklin discovered the double-helix structure of DNA
C. The genetic algorithm was conceived
4. Which came first?
Take a moment to mentally order these chronologically
C. The genetic algorithm was conceived
(1950 — Alan Turing)
A. David Hasselhoff was born
(1952 — Baltimore)
B. Rosalind Franklin discovered the double-helix structure of DNA
(1953 — London)
5. Which came first?
Take a moment to mentally order these chronologically
A. Deep learning neural networks were invented
B. The last people on the moon depart
C. Led Zeppelin IV releases
6. Which came first?
Take a moment to mentally order these chronologically
A. Deep learning neural networks were invented
(October 1971 — 8-layer GMDH, Alexey Ivakhnenko)
C. Led Zeppelin IV releases
(November 1971 — Atlantic Records)
B. The last people on the moon depart
(December 1972 — Apollo 17)
7. Data science is older than you think
The tools and methods have existed for decades.
The technology to leverage at scale them has not.
1960 19661954 1972 1978 1984
Reinforceme
nt learning
for AI studied
First
appearance
of the term
“Deep
Learning”
First
Evolutionary
Algorithm
Implementati
on
Back
propagation
invented for
control
systems
Oracle V1
Implementati
on
8. “[Parameter estimation algorithms] for achieving the minimum in (10),
and the statistical properties of the type (15) are all of general character
and well known for the more traditional model structures used.
They have typically been reinvented and rediscovered in the NN
literature and been given different names there. This certainly has had
an alienating effect on the "traditional estimation" communities.”
- J. Sjöberg, H. Hjalmarsson, L. Ljung, 1994
10. Physical sectors are full of potential
McKinsey’s 2015 MGI Digital America study shows what industries have
the lowest level of digitalization.
● Study was done pre-IoT
● Study was done pre-Data Science revolution
Agriculture and Hunting
Mining
Construction
Health Care
Basic Manufacturing
Chemicals and
Pharmaceuticals
Transportation and
Warehousing
11. Boeing 787 Dreamliner
● 500 GB per flight
● 554 airframes in service
~227 TB fleetwide, per flight!
(That’s very roughly as much as
YouTube generates daily, ish)
14. Data to can save the world
Climate change is real. We cannot afford to wait
for generational improvements to efficiency.
Cycle time for 1% gas turbine efficiency
improvement: 18-120 months.
1% efficiency = millions of dollars in fuel
1% efficiency = millions of tons of CO2
1% efficiency = millions of cars off the road
16. How data are used at various scales
<=100 ms
10 s
1 min
1 hr
1 day
1 week
1 month
1 yr
10+ yr
1 unit 100 units 100000 units
Real-time control systems
Operator feedback
Operational History (single run)
Operational Lifecycle
Real-time fleet monitoring
Operational Logistics
Fleet maintenance scheduling
Service offering/product
development
Fleet purchasing/replacement
Telematics
Maintenance planning, recall
Product development &
engineering
Aggregate over time
Aggregate over units
18. Dynamical Systems Modeling
u x
A dynamical system is a state-space model where the state x changes with respect to time
and in response to the input u. Sometimes we call the model the plant process.
A Linear Time Invariant (LTI) system is the most fundamental dynamical system.
It can be modeled with a first order system of differential equations.
19. Dynamical Systems Modeling
𝚺
e xr
+
-
We can affect the behavior of the system by measuring its state and creating a feedback
loop.
This allows us to set a reference signal r, which we compare to the state to obtain an error e,
which we seek to minimize.
20. I
D
e
Dynamical Systems Modeling
𝚺
xr
+
-
A slightly more sophisticated approach is the Proportional-Integral-Derivative or PID
controller.
The PID controller uses a set of gains, KP, KI, KD that act like weights.
P
𝚺
+
+
+
u
27. Handling Noise with Kalman Filtering
Use state estimate at step k−1 to
predict state at step k
Predict error covariance using
process noise covariance
Compute the measurement error
and its covariance, using
measurement noise covariance
Compute the optimal Kalman gain
Update state estimate,
posterior covariance, and
posterior residual
29. Handling Noise with Kalman Filtering
A Kalman filter is more or less just a Hidden Markov Model!
30. Control Theory Everywhere
We’re building feedback systems all the time, but
how can we design intentionally for them?
● Distributed Systems Monitoring We create
feedback loops with monitoring tools; constrained
Kalman filtering techniques can be leveraged to
create probabilistic real-time anomaly detection.
● Predictive Maintenance As components wear, error
accumulates, which manifests as increased control
demand. This is a useful feature for a prediction
method.
31. Modern machine
learning algorithms
and control theory
have the same
pedigree.
Well-known
Algorithms
We’re generating
immense amounts of
data and doing
almost nothing with
it!
The business case
cannot be more
clear: we can use
data at many scales
to make our systems
and processes
smarter.
We’ve never had
better tools to
implement these
techniques, and they
can be extended.
Tons of
Data
Clear
Business
Value
Existing
Tools
Data Science 💞 Physical Computing
32. “The fruit [of Control Theory] is so low it is
TOUCHING THE GROUND!”
- Colm MacCárthaigh