Machine learning

for calculating tariff prices

Why a techtalk?

by Andrew Ng

5 courses

~80 hours of videos + coding

30 euro / month

The problem

?

?

?

?

?

The problem

3rd party service

  • avg 5 seconds
  • in 20% cases > 7 seconds
  • sometimes > 60 seconds

The solution

  1. Collect the data
  2. Take a single tariff
  3. Use it to train the ai to tell:
    1. Is it in the list (yes/no) ?
    2. What is net price (decimal)
  4. Hope the ai solves the problem

3 months, 331726 entries

ANN - Artificial neural network

 

- Warren McCulloch and Walter Pitts (1943) -

  created a computational model for neural networks

 

- Ramón y Cajal, Santiago (1899).

  Comparative study of the sensory areas of the human cortex.

 

- Frank Rosenblatt build Mark I Perceptron in 1960

 

ANN - Artificial neural network

 

- Universal approximation theorem

  (proved by  George Cybenko in 1989)

 

ANN - Forward propagation

Supervised learning,

error backward propagation

A sample training example

{
  "input": {
    
    "familyStatus": "Keine Angabe",
    "educationType": "Berufsausbildung",
    "jobSituation": "Angestellt/Selbstständig",
    "industry": "Sonstige Branche",
    "benefitAgeLimit": 60,
    "benefitAmount": 1000,
    "fractionOfficeWork": 100,
    "staffResponsibility": 0,
    "smoker": false,
    "birthday": "1983-11-07",
    "insuranceStart": "2018-12-01"
  },
  "output": {
    "exists": true,
    "net": 92.05
  }
}

Enum, 6500 options, actually ~ 6 risk groups

Enum, 4 options

Enum, 9 options

Enum, 3 options

Enum, 10 options

Number

Number

Number

Number

Boolean

Date

Date

Boolean

Number

"occupation": "Isolierer,in",

The input layer

Input layer

single node,

a real number 0..1

 

(the presence problem)

41 input nodes

Hidden

layer

Output

layer

This was easy!

  • Is it a professional approach? - how to tune, where to run, how long to wait?
  • Single occupation - try a bottleneck layer?
  • Max error still huge - try another loss? another learning algorithm?
  • "Predictability" - will it always return "sane" results?

Questions remaining

Made with Slides.com