introduction to machine learning for building design and

Join the conversation #AU2017Join the conversation #AU2017

Introduction to Machine Learning for Building Design and ConstructionMehdi Nourbakhsh, Ph.D.Sr. Research Scientist, Autodesk Researchmehdi.nourbakhsh [at] Autodesk.com

What is AI/ML?

History of ML

Application of ML in AEC

ML Glossary

ML Types

ML Process

ML Algorithms

Decision Tree

Artificial Neural Network

Wrap up

Contents:

What is AI/ML?

History of AI/ML

A lot more here

Day-to-day ML Applications

Demo

Application of ML in AEC

Data, sample

Variables, attributes, features

Feature vector

Input, output

ML Glossary

outlook temperature humidity windy construction

Day 1 overcast hot high FALSE yes

Day 2 rainy cool normal TRUE yes

Day 3 sunny mild high FALSE no

ML Glossary

Feature Vector= (overcast, hot, high, FALSE)

What is the goal?

ML Glossary

Model

Outlook

Temperature

Humidity

Windy

Construction

Input Output or label

F(Outlook, Temperature, Humidity, Windy) = Y

A1 A2 A3 Y

1 0.2 Yes 0.32

5 0.56 No 0.58

2 0.56 Yes 0.23

3 0.6 Yes 0.39

Exercise 1

In the following example, what is/are:• Samples• Features• Input• output

ML Types

A1 A2 A3 Y

1 0.2 Yes 0.32

5 0.56 No 0.58

2 0.56 Yes 0.23

3 0.6 Yes 0.39

Exercise 1.5

Is this a supervised or unsupervised learning?

Is this a classification or regression problem?

Donten in a company that has more than 10,000 building models in their

database. These models are either residential, office space, or warehouses.

They want to categorize their models into these three groups and seek your

recommendation as a thought-leader in the industry. Which one do you

recommend?

Hiring 20 people to manually open all of the models and categorize them in 10

months?

Using machine learning to automatically categorize the models in a month?

(why does it take a month?)

Exercise 2

Which type of ML is this?

How should we describe this

problem to the machine?

Data?

Features?

Input, output?

Exercise 2

Model

?

?…

?

?

Input Output

Residential Warehouse Office

Simplified ML process

In practice, this is not a straightforward process, there are lots of back and forth between the steps.

ML Algorithms: Decision Tree

Predict if construction will happen?

Exercise 3: Decision Tree

Take a guess…


Follow this algorithm:• Divide into subsets• Are they pure? (all

‘yes’ or ‘no’)• Yes->done• No-> repeat

Exercise 3: Decision Tree9 Yes / 4 No

Follow this algorithm :• Divide into subsets• Are they pure? (all

‘yes’ or ‘no’)• Yes->done• No-> repeat

Demo in Weka

Section break (5 min)

ML Algorithms: Artificial Neural Network

Source: https://www.intechopen.com/source/html/39067/media/image1.png

A) human neuron; B) artificial neuron;C) biological synapse; D) ANN synapses

A construction company completed three projects. The estimated and actual

construction cost of each project is presented in the following table. Your task

is to design a cost model for the company based on the available data and

predict how much is the actual cost of ‘proj 4’.

Exercise 4: Construction Cost Model

Estimated Construction Cost Actual Construction Cost

proj 1 350 420

proj 2 200 250

proj 3 500 700

Proj 4 400 ?

A construction company completed three projects. The estimated and actual

construction cost of each project is presented in the following table. Your task is

to design a cost model for the company based on the available data.


Input(X)

Output

(Y)* W


proj 1 350 420

proj 2 200 250

proj 3 500 700

Proj 4 400 ?

One way of doing it:

A= Sum all estimated costs

B = Sum all actual costs

W = B/A



proj 1 350 420

proj 2 200 250

proj 3 500 700

Sum 1050 1370

B/A 1.30

Is it good? Why not 1.5 or 2.0?

Y=1.30X

Measure of goodness = less error

What is the error of the model?


Estimated Construction Cost Actual Construction Cost Predicted Cost (y=1.3x) Error

proj 1 350 420 455 ?

proj 2 200 250 260 ?

proj 3 500 700 650 ?


Absolute error = | predicted cost – actual cost |

Estimated Construction Cost Actual Construction Cost Predicted Cost (y=1.3x) Error

proj 1 350 420 455 -35

proj 2 200 250 260 -10

proj 3 500 700 650 50

We don’t like negative error

Mean absolute error (MAE) = ∑| predicted cost – actual cost | / no of samples

What is the mean absolute error of the model? 31.67

absolute

Let’s try other Ws:


Estimated Construction Cost Actual Construction Cost y=1.3x y=1.4x y=1.5x

proj 1 350 420 455 588 682.5

proj 2 200 250 260 350 390

proj 3 500 700 650 980 975

31.67 182.67 47.5MAE

Input(X)

Output

(Y)* W=1.3Our Neural Net model:

Let’s see if we can do better:


X Y* W + b

Y = W*x + b

input outputweight bias


W/b -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70

1 177 167 157 147 137 127 117 107 97 87 77 67 57 53 50

1.1 142 132 122 112 102 92 82 72 62 52 42 42 45 48 52

1.2 107 97 87 77 67 57 47 37 33 37 40 43 47 50 53

1.3 72 62 52 42 35 32 28 32 35 38 42 45 48 58 68

1.4 37 33 30 27 23 27 30 33 43 53 63 73 83 93 103

1.5 25 22 18 28 38 48 58 68 78 88 98 108 118 128 138

1.6 33 43 53 63 73 83 93 103 113 123 133 143 153 163 173

1.7 68 78 88 98 108 118 128 138 148 158 168 178 188 198 208

1.8 103 113 123 133 143 153 163 173 183 193 203 213 223 233 243

1.9 138 148 158 168 178 188 198 208 218 228 238 248 258 268 278

2 173 183 193 203 213 223 233 243 253 263 273 283 293 303 313

?

Y=2x-70

Y=x Y=x+70

Y=2x+70

Y=x-70

What is the mean absolute error of these models?


W/b -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70

1 177 167 157 147 137 127 117 107 97 87 77 67 57 53 50

1.1 142 132 122 112 102 92 82 72 62 52 42 42 45 48 52

1.2 107 97 87 77 67 57 47 37 33 37 40 43 47 50 53

1.3 72 62 52 42 35 32 28 32 35 38 42 45 48 58 68

1.4 37 33 30 27 23 27 30 33 43 53 63 73 83 93 103

1.5 25 22 18 28 38 48 58 68 78 88 98 108 118 128 138

1.6 33 43 53 63 73 83 93 103 113 123 133 143 153 163 173

1.7 68 78 88 98 108 118 128 138 148 158 168 178 188 198 208

1.8 103 113 123 133 143 153 163 173 183 193 203 213 223 233 243

1.9 138 148 158 168 178 188 198 208 218 228 238 248 258 268 278

2 173 183 193 203 213 223 233 243 253 263 273 283 293 303 313

What model is better?

W/b -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70

1 177 167 157 147 137 127 117 107 97 87 77 67 57 53 50

1.1 142 132 122 112 102 92 82 72 62 52 42 42 45 48 52

1.2 107 97 87 77 67 57 47 37 33 37 40 43 47 50 53

1.3 72 62 52 42 35 32 28 32 35 38 42 45 48 58 68

1.4 37 33 30 27 23 27 30 33 43 53 63 73 83 93 103

1.5 25 22 18 28 38 48 58 68 78 88 98 108 118 128 138

1.6 33 43 53 63 73 83 93 103 113 123 133 143 153 163 173

1.7 68 78 88 98 108 118 128 138 148 158 168 178 188 198 208

1.8 103 113 123 133 143 153 163 173 183 193 203 213 223 233 243

1.9 138 148 158 168 178 188 198 208 218 228 238 248 258 268 278

2 173 183 193 203 213 223 233 243 253 263 273 283 293 303 313


X Y* W + b

Y = 1.5*x -50

input outputweight bias

Congratulation! You have trained your first neural network model!!!


Y = 1.5*x -50


proj 1 350 420

proj 2 200 250

proj 3 500 700

Proj 4 400 ?

Y = 1.5*400 -50 = 550 answer


X1

Y

*W1

+ b

X2 *W2

Y = W1x1+ W2x2 + b

Estimated Construction Cost Average temperature Actual Construction Cost

proj 1 350 62 420

proj 2 200 55 250

proj 3 500 85 700

Proj 4 400 53 ?


X1

f

*W1

+ b

X2 *W2

f = W1x1+ W2x2 + b

T(f)

Weighted sum of input

Activation function

YPredicted value

Exercise 5:

X1

f

*W1

+ b

X2 *W2

f = W1x1+ W2x2 + b

T(f) Y ?

What is the predicted value (Y) if:X1=0.2 , X2= -0.3W1=0.5, W2=0.5b= 0.1

Exercise 5:

X1

f

*W1

+ b

X2 *W2

f = W1x1+ W2x2 + b = 0.2*0.5-0.3*0.5+0.1=0.05T(f)=T(0.05) = 1

T(f) Y ?

What is the predicted value if:X1=0.2 , X2= -0.3W1=0.5, W2=0.5b= 0.1

Some activation functions

threshold sigmoid

tanhReLU


Now, let’s talk about a group of neurons working together.

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

X1

f

*W1

+ b

X2*W

2

T(f) Y

Hmmm… looks very messy. Let’s come up with a shorter from of a neuron.

ML Algorithms: Shorter form of a neuron

X1

T(F=∑(XiWi+b))

*W1

b

X2 *W2

X1

f

*W1

+ b

X2 *W2

T(f) Y

ML Algorithms: Short form of a neuron

n

X1

T(F=∑(XiWi+b))

*W1

b

X2 *W2

�� layer

Neuron at this layer

n

n

n

n

n

n

Can you identify input/output/hidden layer?

Input layer Output layer

Hidden layer

What is the error?

Error = |y- ��|

Demo in Weka

What are the repetitive tasks that you do in your everyday job and ML can help

you to automate?

Exercise 5

What are the problems that you have no idea how to solve or things that you

have no idea how to do, and ML can help you to do or resolve it?

Exercise 5

Imaging you have a personal assistant agent that is like a trusted collaborator: it

can brainstorm with you, it can do research for you, it can tell if your idea

doesn’t work because of this and that, it can fill in the role of other parties (if

you are an architect, it can provide you with the knowledge of engineering or

construction). What are the things that you want your assistant to do for you.

Exercise 5

Wrap up

What is AI/ML?

History of ML

Application of ML

ML Glossary

ML Types

ML Process

ML Algorithms

Wrap up

What is AI/ML?

History of ML

Application of ML

ML Glossary

ML Types

ML Process

ML Algorithms

Model

Outlook

Temperature

Humidity

Windy

Construction

Input Output or label

F(Outlook, Temperature, Humidity, Windy) = Y

Wrap up

What is AI/ML?

History of ML

Application of ML

ML Glossary

ML Types

ML Process

ML Algorithms

Wrap up

What is AI/ML?

History of ML

Application of ML

ML Glossary

ML Types

ML Process

ML Algorithms

Decision Tree

Neural Network

Autodesk and the Autodesk logo are registered trademarks or trademarks of Autodesk, Inc., and/or its subsidiaries and/or affiliates in the USA and/or other countries. All other brand names, product names, or trademarks belong to their respective holders. Autodesk reserves the right to alter product and services offerings, and specifications and pricing at any time without notice, and is not responsible for typographical or graphical errors that may appear in this document.

© 2017 Autodesk. All rights reserved.

introduction to machine learning for building design and

Documents